The Orchestration Economics Manifesto

The Surface Cracked Open

The software industry’s 2025 optimism masked structural transformation. When it surfaced in early 2026, the market had no framework to interpret it.

The software industry entered 2024 in what looked like a renaissance. AI launches pushed SaaS valuations higher^{13Bessemer Venture Partners, “State of the Cloud 2024.” Software Equity Group, “Public SaaS Company Market Valuations” / SaaS Index Q4 2024. SaaS Capital, “The SaaS Capital Index.” Aventis Advisors, “SaaS Valuation Multiples: 2015–2026.”}. ServiceTitan’s IPO succeeded^{14ServiceTitan, "ServiceTitan Announces Pricing of Initial Public Offering," December 11, 2024}. Salesforce announced Agentforce^{15Salesforce, “Salesforce Unveils Agentforce–What AI Was Meant to Be” (Dreamforce 2024 press release, September 12, 2024)}. Investors believed the next growth cycle had begun. Beneath the optimism, something else was happening.

Over the course of early 2026, the surface cracked open. ServiceNow, Salesforce, Workday, and Adobe each lost between 6% and 15%^{16Market data from February 2 to February 6, 2026} of market capitalization in a single week of trading in February 2026
[Figure 3].

The software industry that began 2025 celebrating AI-driven growth ended the year confronting the narrative that AI-driven growth reflected in their top line mattered less than the perceived vulnerability to their fundamental business model.

The software companies and investors hoped it was merely a correction. But it was a signal. Understanding the signal requires an important distinction.

§

Discontinuity is not Disruption

Disruption steepens or flattens an existing curve. Discontinuity stops it and starts a new one. The absence of mental models is a defining characteristic.

The era of generative and agentic AI is not disruption. It is discontinuity.

I know it is tempting to dismiss these as buzzwords or shallow consultant shorthand. Disruption. Discontinuity. Paradigm. Several decades of overuse have stretched and flattened such words, robbing them of meaning. But these words once signified very precise concepts that shaped how we understood technological change. That is why, when I began writing three years ago about generative and then agentic AI, I chose the word discontinuity. Not because it sounds dramatic. But because it is technically accurate, and the distinction between disruption and discontinuity is load-bearing for everything that follows. I believe we must reclaim that meaning to help navigate the present and guide us toward the future as it unfolds.

Disruption follows predictable patterns. New entrants target overlooked segments, gradually moving upmarket until incumbents fall. An existing curve steepens or flattens, but the curve continues. Leaders and investors can deploy existing tools to analyze the shift. This was the foundational insight of Clayton Christensen’s Innovator’s Dilemma. He first coined the phrase in 1995^{17Bower, J.L. & Christensen, C.M. (1995). "Disruptive Technologies: Catching the Wave." Harvard Business Review}, one year after the Netscape browser^{18EBSCO Research Starters, Release of Netscape Navigator 1.0, December 15, 1994} was released, and his framing for thinking about technological disruption became a cornerstone for the next three decades of innovation and entrepreneurship.

Discontinuity goes beyond disruption. If disruption changes the slope of a curve, Discontinuity breaks the curve entirely. All previous strategies, tools, and historical lessons are ripped away. The break can lead to extraordinary value creation or extraordinary destruction.

When cloud computing disrupted on-premises software, it was a radical change in terms of pricing and technology. And yet, it was still CRM or payroll. People using it and installing it had reference points from their experiences. That is disruption.

Generative and agentic AI broke this pattern. Companies are not being handed a new CRM or payroll platform. Someone is giving them a magic wand and saying, “Here, it can do anything. Imagine what you can do with this”. Nothing in their experience has prepared them for the magic-wand scenario. They have no mental models.

So, they turn back to the old ones and stick the magic wand into the old tech stack, wishing for increased ROI that never materializes. Only 5% of enterprises reported any ROI from their initial generative AI projects last summer when MIT conducted a first-of-a-kind survey^{19Fortune, MIT report: 95% of generative AI pilots at companies are failing, August 18, 2025}. As the study’s authors noted, the failure was not “driven by model quality or regulation but seems to be determined by approach”.

This is not surprising. Most investors and executives optimize for disruption because they can deploy familiar analytical frameworks, frameworks that project cash flows, calculate returns, and assess competitive dynamics using tools refined over decades. Discontinuity places them in a landscape where everything they once knew has lost its power to guide them.

French philosopher Gaston Bachelard described “epistemological rupture” as moments when progress requires abandoning old frameworks entirely, because accumulated knowledge has become an obstacle to understanding.

This is the core difficulty. The greatest obstacle to understanding discontinuity is not the absence of new information. It is the presence of old frameworks that interpret new evidence as confirmation of the regime that is ending.

This discomfort produces two failure modes. The first is paralysis: waiting for clarity that will arrive too late. The second is pattern-matching to previous cycles: applying frameworks that are no longer relevant and making catastrophic allocation decisions.

Frustration grows because there has been a failure to recognize the discontinuity. Meanwhile, the capability of technology is rapidly advancing. Those who have recognized the discontinuity are somehow, almost inexplicably to others, transforming their businesses at unimaginable velocity.

This inability to build new mental models is a defining characteristic of discontinuity. Those unable to re-orient themselves are at grave risk of being on the wrong side of a historic economic valuation shift.

§

The Human Challenge and the Scale of Time

Discontinuity is not only a technological event. It is a cognitive one. People’s brains are not framed for it. And it plays out over timescales that defy the rhythms of capital markets.

This is about tech. But it’s a mistake to think that it’s just about tech. Such moments present a fundamental human challenge: People’s brains are not framed for discontinuity.

Navigating discontinuity requires an abstraction capability, the ability to see how all the pieces might eventually align and what consequences could be unleashed. Generative and agentic AI have the power to turn many non-technology companies into technology companies, an evolution many industries should start recognizing now.

To understand and adapt to discontinuity requires an understanding of the underlying components that are driving this phenomenon, having a clear vision on how that impacts an existing business model, and developing a sophisticated view around all the structures and assets that are realigning or emerging to enable something new.

This has been at the very core of my work for the past three years. And if there is one critical aspect that must be recognized to begin building the right mental model for this discontinuity, it is understanding the scale of time. The generative and agentic AI transformation is far more profound and sweeping than even many sophisticated technology leaders have truly grasped. That means it will play out over a much longer time frame than investors and entrepreneurs are accustomed to tolerating. This shift requires everyone across the innovation economy to fundamentally reset the way they think about building new companies, making investments, and defining the shape of markets. This moment demands a long-term view to map out which investments and technologies will make sense at which moment and why.

Amid the rubble of the dot-com bust, a certain defeatism set in. Maybe all the internet hype had really been just that: hype. But many Silicon Valley VCs began discussing a book written a decade earlier by Science Historian David Nye called, “Electrifying America: Social Meanings of a New Technology” ^{20MIT Press, Electrifying America, July 8, 1992}. The book traces the vast ways that electricity transformed the nation.

Cities installed electric streetlights, which expanded nightlife and caused a wave of new restaurants to open. To get more customers to demand electricity in their homes, they designed electrical appliances like refrigerators and washing machines. No one could have envisioned most of these innovations when Thomas Edison made his first lightbulb in 1879. It took decades to put the infrastructure in place, raise the investment capital, develop the use cases, and refine the business models.

The real transformation played out over a much longer timescale as all the pieces were put into place. Such was the case with digital transformation. Though it may have taken longer than initially predicted, today we are living in that digital world many envisioned in 2000 thanks to smartphones, 4G wireless, broadband, and a host of business model innovations enabled by these waves of infrastructure.

When dealing with discontinuity, the scale of change is so massive that timelines are impossible to predict. So many pieces must align before exponential potential is unleashed.

Then, when it happens, the speed of change is startling.

This pattern repeats. A breakthrough technology appears. Euphoric predictions about sweeping changes are made. They fail to manifest. Doubt appears. Progress seems to stall. And then, in a blink, everything accelerates with little warning. We are in that phase now.

Anyone in the tech industry who experienced 2025 felt every extreme of this cycle. One moment, there was anxiety about an AI bubble. Then there was fear of overspending on infrastructure. Then the pendulum swung, and the market believed a niche product announcement crushed billions of dollars in market value for incumbent SaaS companies. Now, in reverse, we are all panicked that we are running out of compute and assessing whether compute constraints constitute or not a single point of failure (SPOF) for the AI labs.

OpenAI CEO Sam Altman captured the difficulty of reconciling these extremes when he said in August 2025: “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes. Is AI the most important thing to happen in a very long time? My opinion is also yes” ^{21CNBC, OpenAI’s Sam Altman sees AI bubble forming as industry spending surges, August 18, 2025}.

The bubble skeptics are correct that many companies in this space will fail, just as many internet companies failed in 2000. They are wrong to conclude that the transition itself is speculative.

§

The Capability Curve Broke First

What was impossible became possible overnight. The capability is proven and reproducible across multiple vendors.

This discontinuity creates the foundation for something new: abundant intelligence, accessible through APIs, improving exponentially.

Companies that relied on linear improvement in AI capabilities, expecting gradual progress from models that could barely complete sentences to models that could write paragraphs now suddenly faced step-function changes that break all projection curves.

This is not about doing existing things 10% better. When ChatGPT was first released in November 20226, AI could describe a concept. By 2025, AI could architect a solution, write the code, build the interface, and deploy a functional application autonomously. By March 2026, AI could coordinate teams of sub-agents that divide complex professional work across domains, execute in parallel, sustain reasoning across a million tokens of context, operate computers through their own interfaces, and correct their own errors over hours of unsupervised operation. Three years. Three different worlds.

Importantly, this capability was proven and reproducible across multiple vendors. It was not one model, one company, or one approach. Frontier performance converged: multiple models achieve 80% on SWE-Bench (Software Engineering Benchmark) [Figure 5] that evaluates large language models by testing their performance on real-world software engineering tasks. Multiple coding tools are capturing billions in revenue. The leaderboard positions for the latest frontier models change from day to day, week to week. The progression is accelerating in a way fundamentally different from previous AI winters.

The gap between capability announcement and production deployment has been compressed from years to months. GPT-4’s initial capabilities^{22OpenAI, "GPT-4 Technical Report," March 14, 2023} rolled out to a limited set of users over several months in 2023 with tight restrictions, and then gradually expanded to more general availability over the next year. In sharp contrast, when Google released Gemini 3 Pro in November 2025^{23Google, A note from Google and Alphabet CEO Sundar Pichai, November 18, 2025}, it was available globally, already baked right into many of its most widely used products. What once required a year of cautious rollout now ships in a morning.

Models are reaching PhD-level reasoning as a foundational capability. Not as research artifact, but as commercially available infrastructure.

§

Then the Substrate Commoditized

PhD-level reasoning went from nonexistent to commercially available to commodity-priced in eighteen months.

As AI capabilities advanced rapidly, a group of researchers s in late 2023 proposed the graduate-level Google-Proof Q&A (GPQA) benchmark, based on 448 PhD-level questions from multiple disciplines. A more rigorous variation based on a subset of 198 questions emerged, known as GPQA Diamond, where human experts scored 69.7%^{24David Rein et al., "GPQA: A Graduate-Level Google-Proof Q&A Benchmark", arXiv:2311.12022, November 20, 2023}. The first commercially available model did not surpass that human threshold until December 2024. Eighteen months later, multiple models from competing laboratories exceeded 90% [Figure 6].

Meanwhile, the cost per query had fallen to fractions of a cent. Peer-reviewed research from Epoch AI found that the price to achieve a given level of frontier performance was declining by a factor of 5-10x per year^{25Hans Gundlach et al., "The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference", arXiv:2511.23455, November 2025}. Roughly a third of that decline was attributable to algorithmic efficiency alone.

The convergence was as significant as the cost trajectory. On the Massive Multitask Language Understanding (MMLU) benchmark, the standard measure of general knowledge across academic subjects created in 2020, the gap between proprietary frontier models and their open-source alternatives collapsed from approximately 15 percentage points to less than one by late 2025 [Figure 8].

Each month brought a new release that narrowed the distance further. DeepSeek, Kimi K2, and Llama achieved near-parity with systems that cost orders of magnitude more to build. Intelligence that was once the exclusive province of three or four laboratories became available to anyone with an API key or a sufficiently capable laptop. When the substrate becomes uniform, differentiation migrates to what is built upon it. This is not a theoretical observation. It is the mechanism that drives everything that follows.

Abundant intelligence is the necessary condition. The discontinuity is what that abundance made possible: machines that act.

§

Then Machines Became Actors

Abundant intelligence was the necessary condition. It was not the break.

The break occurred when intelligence became sufficient for autonomous action, when machines crossed from processing instructions to pursuing goals. This was a qualitative change in what software is. Every previous generation of software, including every previous generation of AI, operated within the same architectural relationship to human work: the human decided, the tool executed.

The tool could be astonishingly capable. AlphaGo defeated the world champion at Go ten years ago^{26DeepMind, "AlphaGo: Mastering the ancient game of Go with Machine Learning," Google DeepMind Blog, 2016}. It did not decide to play Go. GPT-4 could draft a legal brief. It did not decide which brief to draft, did not select the relevant precedents from the firm’s case history, did not coordinate with the associate handling discovery, did not file the brief with the court. The human performed every coordination function. The AI performed one step, brilliantly, and waited for the next instruction.

Between late 2024 and early 2026, that architecture broke.

The evidence is not ambiguous. When Claude Code receives a task such as “resolve this GitHub issue”, it reads the codebase, identifies the relevant files, reasons about the bug, writes a patch, runs the test suite, discovers the patch introduced a regression, rewrites the patch, reruns the tests, and submits a pull request. No human selected the files. No human ran the tests. No human identified the regression. The agent received a goal and autonomously determined how to achieve it.

That is not a better tool. That is a new kind of actor.

The distinction matters because it determines the correct analytical framework. If machines are better tools, the old frameworks apply with modest adjustment: productivity improves, margins expand, the curve steepens. The analyst’s DCF model works. The board’s strategic plan holds. The investment bank’s sector thesis survives with updated assumptions.

If machines are actors, the frameworks break. The unit of production changes. The cost structure of work changes. The competitive dynamics of every industry that employs knowledge workers change. The analyst needs a new model, not updated assumptions.

That is what Kuhn meant by a paradigm shift. For those that still associate “paradigm shift” with Kuhn, the phrase can almost conjure a mental image of gliding from one framework to another. What is often overlooked is an equally important part of Kuhn’s insight in terms of what catalyzed these shifts. In the scientific communities, an existing paradigm would begin to experience a series of “anomalies”. These were observations the framework could not explain, plunging it into a period of crisis.

The “paradigm shift” often involved a messy, heated middle period full of disagreement over direction, rules, and ways to explain and interpret the new information and dynamics. Many schools of competing thought would emerge trying to define the rules of the new paradigm, a necessary, if sometimes brutal process. While Kuhn spoke of “anomalies” in this, we have chosen a different, precise word to guide us: discontinuity.

To navigate through discontinuity and into the new paradigm, it is important to recognize that the old observations are wrong. They are incommensurable. The old framework cannot produce the new observations. The data that would explain the shift does not fit into any field the old model contains.

Consider the shock when software companies lost hundreds of billions in value in early February 2026. Entire sectors were repriced on the assumption that something fundamental had changed. They were right. But the shift was not caused by a single model release or a single company. It emerged from a cascade of changes across the AI stack in cost, capability, infrastructure, coordination, and deployment that arrived within months of one another. Each was significant alone.

Together they produced the first structural signal of the Agentic Era: intelligence behaving like a commodity. When intelligence commoditizes, value does not disappear. It migrates.

But to where, to whom, and on what terms?

Those questions are the core of this manifesto. Coding was the first domain to cross, but its lessons generalize to any domain involving structured reasoning, tool use, and verifiable outputs. The crossing order is predictable. The crossing itself is not in question.

Chapter 2 traces the cascade that forced this reckoning into the open.

A single shock disrupts. A cascade of shocks, each in a different domain, each amplifying the others, forms a discontinuity and ushers us in a new paradigm.

The theory of “punctuated equilibrium”, popularized by evolutionary biologist Stephen Jay Gould, posits that long periods of stability are occasionally interrupted by rapid change. Significant evolutionary leaps happen quickly rather than steadily accumulating over time. Species appear stable for millennia, then shift in geological instants. Gould’s deeper insight was that individual punctuations do not matter in isolation. What matters is when multiple punctuations occur within the same geological window and reinforce each other.

Between September 2024 and February 2026, six rapid phase shifts arrived across different layers of the AI stack:

Intelligence arrives: OpenAI’s o1 crossed the PhD reasoning threshold. The capability existed for the first time^{27OpenAI, "Learning to Reason with LLMs" (o1 announcement), September 2024}.

Intelligence becomes accessible: DeepSeek and open-source models proved that frontier capability no longer required frontier budgets^{28DeepSeek-R1 release, January 2025; see analysis in Stanford HAI, 2025 AI Index Report}.

Silicon independence: Gemini 3 Pro demonstrated frontier performance on non-NVIDIA chips^{29Intuition Labs, Google TPUs Explained: Architecture & Performance for Gemini 3, November 2025}.

Protocol standardization: Model Context Protocol (“MCP”) became the universal coordination layer, surpassing 97 million monthly downloads^{30Model Context Protocol Blog, "MCP joins the Agentic AI Foundation", December 9, 2025}.

Inference swarm: Moltbook’s agent swarm generated millions in daily compute spending from autonomous agents operating without human attention^{31Yukun et al., “Humans welcome to observe”: A First Look at the Agent Social Network Moltbook, February 2026}.

Long-context frontier: The one-million-token context window crossed from research capability to production infrastructure, enabling agents to reason across entire codebases and filing suites^{32Google Gemini 3 Pro / 1M-token context window in production, November 2025. See IntuitionLabs, "Google TPUs Explained: Architecture & Performance for Gemini 3", December 2025}.

They laid the structural foundation for the Agentic Era in which intelligence is cheap, ubiquitous, and coordinated at machine scale. Intelligence is no longer the scarce resource on which competitive advantage can rest

Each was a tremor. Together they traced one fault line. This chapter follows the crack.

§

Shift 1 – Intelligence Arrives

In September 2024, AI crossed the PhD threshold. The capability that would reshape the economy existed for the first time.

For years, frontier models impressed without threatening. They could draft, summarize, and generate code. These were useful augmentations that left the structure of professional work intact. GPT-4, the most capable model available through most of 2023 and until mid-2024, scored 39%^{33Rein et al., "GPQA: A Graduate-Level Google-Proof Q&A Benchmark," arXiv:2311.12022, November 2023} on GPQA Diamond. The machine was fluent. It was not expert.

Google DeepMind’s Gemini 1.5, unveiled in February 2024^{34Google, "Introducing Gemini 1.5, Google’s next-generation AI model," February 15, 2024}, demonstrated significant advances in multi-step reasoning. Anthropic’s Claude 3 Opus followed in March 2024^{35Anthropic, "Introducing the next generation of Claude," March 4, 2024}, setting new benchmarks in complex analytical tasks. OpenAI’s Strawberry model arrived in September 2024^{36OpenAI, "Learning to Reason with LLMs," September 12, 2024}, released as “o1”, pushing reasoning capabilities further still. All three relied on massive model architectures and extensive computational resources. This created a soaring market for the companies that provided the infrastructure, most notably placing chipmaker NVIDIA on a trajectory to becoming the world’s most valuable company^{37Statista / Bloomberg, June 2024; Washington Post, November 5, 2024} and a kind of AI kingmaker.

But o1 did something the others had not. It scored 77.3% on GPQA Diamond, making it the first model to surpass PhD-level human performance on a benchmark designed to be their floor. The model did not merely generate plausible-sounding answers faster. It reasoned: decomposing problems, testing hypotheses, and revising its approach when intermediate steps failed.

This was a qualitative shift in what machine intelligence could do.

Claude 3.5 Sonnet, Gemini 2.0, and a succession of reasoning models arrived in rapid sequence, each pushing scores higher and prices lower. By early 2026, multiple models from competing laboratories exceeded 90% on GPQA Diamond [Figure 7], surpassing not just the PhD threshold but the PhD experts themselves. The capability had been proven and replicated by at least five independent sources.

The first foundation stone of the agentic paradigm was laid: the intelligence required to perform expert-level cognitive work existed as a commercially available capability. What remained was access, coordination, and then letting it loose.

§

Shift 2 – Intelligence Becomes Accessible

DeepSeek shattered the assumption that frontier intelligence required frontier budgets. Within months, near-parity models were available at commodity pricing.

The arrival of PhD-level reasoning would have mattered less if it had remained the exclusive province of three or four well-capitalized laboratories charging premium prices and able to raise billions for compute infrastructure. A Chinese entrepreneur shattered the assumption that it might in early 2025.

Liang Wenfeng had founded DeepSeek about two years earlier in 2023^{38Fortune, "Meet the hedge fund manager who founded DeepSeek," January 27, 2025}, leveraging his background in quantitative finance to pursue software-driven efficiency rather than brute computational scale. The company embraced open source and a fundamentally different philosophy of model building^{39DeepSeek-AI, "DeepSeek-V3 Technical Report," December 2024, arXiv:2412.19437} that challenged the entire economic calculus of building large language models.

In January 2025, the company released DeepSeek-R1, which matched ChatGPT’s performance on several benchmarks while costing 20-50x less^{40Sinéad Carew, Amanda Cooper et Ankur Banerjee, "DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss", Reuters, 28 January 2025} to run than OpenAI’s o1, depending on the workload/tasks. The market reaction was immediate. The Nasdaq fell more than 600 points. NVIDIA lost nearly $600 billion in market value in a single day36, the largest one-day loss in history. Broadcom shed $200 billion.

The debate over whether DeepSeek achieved its efficiency through genuine architectural innovation or through distillation from Western frontier models missed the structural point. Regardless of method, the result was the same: near-frontier intelligence was now available at commodity pricing. And the pattern continued to accelerate. By August 2025^{41DeepSeek-AI, "DeepSeek-V3.1" model card, HuggingFace, August 21, 2025}, DeepSeek-V3.1, a hybrid model that switched between thinking mode for complex reasoning and traditional interactions, delivered roughly 90% of GPT-5’s capability^{42Artificial Analysis, GPT-5 (high) (Intelligence Index 45)} at 1-2% [Figure 12] of the cost, with self-hosting prices as low as $0.01-0.05 per million tokens^{43CloudZero, DeepSeek Pricing Explained, August 2025; DeepSeek API Docs, Models & Pricing} compared with $3.44 for GPT-5 and up to $30 for Claude.

Adoption followed quickly. Andreessen Horowitz partner Martin Casado noted in August 2025 that roughly one in five startups pitching the firm with open-source models were already using Chinese models^{44Martin Casado, clarification post on X; original quote in The Economist, August 2025. For the corrected figure, see: OfficeChai, "80% of Startups Using Open-source Models Are Using Chinese Models}, suggesting that the cost advantage was no longer theoretical. When OpenAI released GPT-5 a few weeks later39, the benchmarks looked strong, scoring 74.9% on SWE-bench Verified and 94.6% on AIME 2025^{45OpenAI, "Introducing GPT-5," August 7, 2025}.

In November 2025, open and private models converged further^{46Hugging Face, moonshotai/Kimi-K2-Thinking, November 2025; Moonshot AI 2025 Journey, Kimi Official}. Moonshot AI released Kimi K2 Thinking, an open-source model that achieved state-of-the-art performance on Humanity’s Last Exam at 44.9%^{47UNU Campus Computing Centre, The Agent Has Landed: Why Kimi K2 Thinking Is More Than Just a New AI Model, December 2025; Artificial Analysis, HLE Benchmark Leaderboard, April 2026}, surpassing both GPT-5’s 41.7% and Claude Sonnet 4.5, at a reported $4.6 million in training costs for its trillion-parameter architecture. For context: This represented a ~15-40x cost reduction relative to 2023 frontier training runs ($78M for GPT-4, $191M for Gemini Ultra) and was comparable to GPT-3’s reported 2020 training cost, now producing a reasoning system that outperformed the most sophisticated proprietary models on a benchmark specifically designed to be unsolvable by current AI.

$4.6 million. The number deserves repeating because of its implications.

The innovation went beyond cost reduction. Kimi K2 employed what Moonshot called “interleaved thinking and tool use” ^{48DeepLearning.AI, The Batch, "Kimi K2 Thinking Outperforms Proprietary Models With New Techniques for Agentic Tool Use," November 19, 2025}. This methodology uses reasoning tokens and function calls to alternate fluidly within the same inference pass. The model thinks, acts, observes results, thinks again, acts differently based on new information, and continues this dynamic cycle for hundreds of steps without degradation.

Unlike DeepSeek, the detonation failed to shake markets or rattle executive nerves. It went largely unnoticed. But the signal was unmistakable: open source had caught up to the closed-model frontier, not just in raw capabilities, but in the sophisticated reasoning and agentic behavior that defines the next wave.

Open-source models had achieved near-parity with the world’s most expensive systems. Intelligence was no longer scarce. It was no longer expensive. The second foundation stone was laid. Anyone with an API key or a sufficiently capable laptop could access reasoning capability that had not existed at any price eighteen months earlier.

§

Shift 3 – Silicon Independence

Google proved that controlling chips, models, and distribution simultaneously drives costs below what any single-layer competitor can match.

On November 18th, 2025, one week after Kimi K2, Google released Gemini 3 Pro^{49Google, "Gemini 3: Introducing the latest Gemini AI model from Google," November 18, 2025} and demonstrated what happens when the cost collapse meets full-stack control.

The benchmarks were unambiguous: 76.2% on SWE-Bench Verified, 91.9% on GPQA Diamond^{50Benchmark results from respective leaderboards: SWE-Bench (Princeton), GPQA (Google), MMMU-Pro (multi-institution), Aider Polyglot.}, 81% on MMMU-Pro for multimodal understanding, and the highest-ever Elo score on LMArena at 1,501 at the time. Gemini 3 Pro was, by most measures, the strongest frontier model available. The benchmark performance was the headline. What produced it was the real story.

Previous Gemini generations had been trained on Google’s Tensor Processing Units^{51SemiAnalysis, "TPUv7: Google Takes a Swing at the 900lb Gorilla," November 2025}, but they had not topped the industry leaderboards. NVIDIA-trained models had. Gemini 3 broke that pattern. The strongest frontier model available according to SWE-Bench, GPQA Diamond and LLMArena Elo had been trained entirely on non-NVIDIA silicon45. This had never happened before. The company that controls more than 90%^{52Statcounter GlobalStats, February 2026} of global search, 3.5 billion Android active devices^{53Business of Apps, David Curry, "ChatGPT Revenue and Usage Statistics (2026)", updated March 30, 2026}, and the world’s largest cloud infrastructure had demonstrated that frontier AI performance did not require NVIDIA chips.

The consequences extended beyond silicon independence. The competition had moved from a single layer to vertical integration. The debate was not who had the best model. It was: Who controls enough of the stack to optimize the entire chain? The economic signal was as significant as the technical one.

Google could offer Gemini 3 Pro at input pricing 2.5 times cheaper than Claude Opus 4.6 [Figure 13] for comparable performance on standard tasks because Google manufactured its own silicon, trained on its own infrastructure, and distributed through products already used by billions. The cost advantage was an architectural inevitability of vertical integration.

Google also had a massive distribution advantage. It could bundle Gemini 3 into the productivity and cloud stacks already used by hundreds of millions of developers and workers. In a leaked memo reported by The Information, OpenAI CEO Sam Altman had written that Google’s next version of Gemini could “create some temporary economic headwinds for our company” ^{54eWeek, Sam Altman warns OpenAI staff of “economic headwinds” amid Google’s Gemini 3 comeback, November 24, 2025}, adding, “I expect the vibes out there to be rough for a bit”. His assessment proved correct. While Magnificent 7 stocks fell 7.6% between late October and late November, Google rose 18%^{55From 29 October to 25 November 2025. The Magnificent 7 stock decline is calculated as a market-cap-weighted average}. The market was rendering its verdict: full-stack control was worth more than model leadership.

The third foundation stone was in place: the cost of intelligence is driven not only by algorithmic efficiency and open-source competition but by vertical integration of the kind only a handful of companies in the world can execute.

§

Shift 4 – Protocol standardization

MCP became the TCP/IP of the Agentic Era. The coordination substrate went from nonexistent to near-universal in twelve months

If DeepSeek commoditized intelligence, MCP standardized how it would be coordinated.

Anthropic’s Model Context Protocol (MCP) was released in November 2024^{56Anthropic, "Introducing the Model Context Protocol," November 2024. Subsequent adoption data from npm download statistics and GitHub repository metrics.}, providing a standardized way for AI agents to connect to external services, invoke tools, and maintain context across interactions. Within months, it became the industry’s universal connector, adopted even being adopted by competitors like OpenAI and Google. Anthropic donated MCP in December 2025 to the Linux Foundation to form the Agentic AI Foundation^{57The Linux Foundation, Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md, December 9, 2025}. By January 2026, MCP had surpassed 97 million monthly SDK downloads^{58Anthropic, "MCP joins the Agentic AI Foundation," Model Context Protocol Blog, December 9, 2025.}.

What started as a modest product launch became a TCP/IP moment for the Agentic Era. MCP rapidly evolved into the plumbing through which intent flows and context accumulates. The impact of MCP made few headlines and triggered no major stock market panics, but its structural impact was deep. It defined how intelligence would be coordinated. It enabled the deployments of agents in ways that increasingly became cost-effective as users were able to assert more effective control.

Whoever defined the coordination standard could shape the architecture of the entire era that followed. The protocol layer’s technical architecture and competitive implications are examined in detail in Part II. What matters here is the phase shift itself: the coordination substrate moved from nonexistent to near-universal adoption in twelve months.

§

Shift 5 – The Inference Swarm

Moltbook proved that autonomous agents could generate millions in daily compute spending without human attention. Agent populations, unlike human populations, can double in days.

The fifth phase shift arrived in January 2026 with the Moltbook phenomenon^{59TechCrunch, OpenClaw’s AI assistants are now building their own social network, January 30, 2026}, a Reddit-like platform for autonomous agents that drove countless headlines alternating between wonder, humor, and alarm at the implications of machines running amok and talking among themselves.

Look past the theater and content to see the real revelation: autonomous agents generating millions in daily compute spending without human attention. When agents operate at machine speed, producing hundreds of API calls per hour, the demand curve decouples from human work patterns.

There are no evenings, weekends, or vacation days in inference demand from agents. The “demand swarm” is not a metaphor. It is a description of compute load that scales with the agent population rather than the human population. Agent populations, unlike human populations, can double in days^{60Wiz, "Hacking Moltbook: AI Social Network Reveals 1.5M API Keys," February 2, 2026}. Agent demand scales with agent population, not human population.

Moltbook is the description of compute load with no modern precedent.

§

Shift 6 – The Long-Context Frontier

When a model holds an entire codebase in working memory, agentic coding shifts from assisted editing to autonomous engineering.

We have defined the five phase shifts that have created sufficient conditions for the Agentic Era. But even as we move into that new era, the foundational phase shifts will continue. As of April 2026, another phase shift is likely imminent, powerful but also somewhat different: the long-rumored release of DeepSeek-V4^{61DeepSeek V4 Expected to Launch in Late April," April 2026}. The new version will reportedly be a trillion-parameter open-weight model with a one-million-token context window, built on three architectural innovations that are targeting SWE-Bench above 80% at 10-40x below Western frontier costs:

Manifold-Constrained Hyper-Connections

Engram Conditional Memory

Sparse Attention

The long-context dimension is where this matters most. The previous shifts were about capability. This one will be as well. But it will also more clearly redefine advantage. In the Agentic Era, value shifts away from model power to those who control the locus of intelligence.

There is a qualitative threshold within the progression from thousands to millions of tokens. At thousands, an agent reviews a file. At tens of thousands, a module. At hundreds of thousands, a service. At one million with reliable retrieval, an agent holds an entire codebase in working memory. That means every file, dependency, test, and configuration. That allows it to reason across the codebase as a coherent system.

Below this threshold, agentic coding is assisted editing: the model helps a human modify files one window at a time, with the human providing the architectural overview. Above it, agentic coding is autonomous engineering. The model no longer operates on fragments; it understands the system. It refactors across files without losing coherence and traces failures end-to-end without decomposing them into disconnected steps.

DeepSeek-V4’s potential to deliver this at open-weight pricing would democratize autonomous engineering. What was previously confined to frontier labs becomes available to every startup and developer with commodity hardware. The economic threshold at which enterprises deploy coding agents drops precipitously. Adoption no longer flows through enterprise sales cycles, but through developers, compounding from the bottom up.

If DeepSeek-V4 matches Claude Opus 4.6’s million-token capability^{62The one million token context window became generally available for Claude Opus 4.6 (and Sonnet 4.6) at standard pricing on March 13th, 2026.} at open-weight pricing, the long-context frontier commoditizes in the same pattern as reasoning before it. The conclusion is the same: intelligence commoditizes. The value migrates upward.

There is a second-order effect. When long-context models operate over entire codebases, every interaction becomes in-context learning. The model absorbs patterns, conventions, and architectural decisions of the system it inhabits. The agent that has processed a thousand pull requests for a specific repository develops something analogous to institutional knowledge because it was placed inside the system.

This is a change in where learning happens. The system becomes the training ground. Long context accelerates the emergence of systems that learn how to use agents by orders of magnitude.

§

One Fault Line

These shifts must be understood as a single event, not as a sequence of independent developments.

Discontinuity does not move in a straight line. There are moments of acceleration, followed by what feels like a plateau where stagnation has taken hold. Often, in ways big and small, several critical elements align independently of each other, each necessary for the next stage of acceleration.

The phase shifts of 2024, 2025, and early 2026 were the defining examples. The intelligence. The accessibility. The vertical integration. The protocol. The swarm. The long-context frontier. None of these alone would have produced the discontinuity.

Intelligence without accessibility would have remained a premium product for the few. Accessible models without coordination protocols would have been powerful curiosities, but isolated. Protocols without cheap intelligence would have been empty plumbing. Autonomous agents without model convergence would have been too expensive to deploy. Vertical integration without protocol standardization would have fragmented the stack rather than unified it.

Together they produced something that did not previously exist: intelligence that is cheap, ubiquitous, coordinated at machine scale, and accessible to anyone with an API key. This is the structural foundation of the agentic paradigm.

The cost of intelligence will continue to fall. Models will continue to converge. Protocols will continue to standardize. Agent populations will continue to grow. These trajectories already in motion, governed by economics that reinforce rather than exhaust themselves.

But a foundation is not a building. The substrate is in place. Chapter 3 examines the moment it became visible, when models crossed from assisting professionals to performing professional work, and when agents began acting without being asked.

There comes a moment in every exponential curve when the abstract becomes concrete, when the thing everyone has been debating arrives in the specific. For generative and agentic AI, that moment came in the winter of 2025-26. It came in three waves: the models crossed a line, the revenue proved it was real, and the agents began acting on their own.

§

The Models Cross a Line

The intelligence arrives and models cross from assistance to performance.

Until late 2025, frontier AI models were impressive but bounded. They could draft, summarize, translate, and generate code. These tasks were useful. But any honest assessment would classify them as sophisticated assistants, or “copilots”. A lawyer still reviewed every clause the model suggested. An analyst still checked every number. A developer still read every line. The model accelerated human work. It did not perform it.

In the span of twelve weeks between December 2025 and February 2026, that boundary dissolved. OpenAI shipped GPT-5.2 in December 2025^{63OpenAI, "Introducing GPT-5.2," December 11, 2025}, a model whose adaptive reasoning could sustain multi-step analysis across hundreds of pages of financial filings, legal discovery documents, and technical specifications without losing coherence or fabricating connections.

Analysts at early-access firms reported that GPT-5.2 produced equity research drafts that required editing for voice and judgment, not for accuracy or structure. The model was not assisting the analyst. It was performing the analyst’s core function, and the analyst was reviewing output rather than producing it. In February 2026, OpenAI then released Codex^{64OpenAI, "Introducing the Codex app," February 2, 2026 (updated on March 4, 2026)} as a desktop application, positioning it not as a coding assistant but as “command center for agents”, an autonomous worker that receives a task, decomposes it, executes across files, tests its own output, and returns completed work. The framing was deliberate. OpenAI was not selling a better autocomplete. It was selling a junior colleague.

Meanwhile, Anthropic had released Claude Opus 4.6 on February 5, moving the ceiling even higher. A one-million-token context window that genuinely worked: 76% retrieval accuracy on the MRCR v2 needle-in-a-haystack test at full context, compared to 18.5% for Claude Sonnet 4.5^{65Anthropic, "Introducing Claude Opus 4.6," February 5, 2026}. This meant the model could hold an entire codebase, an entire contract suite, an entire quarter’s financial filings in working memory and reason across all of it simultaneously.

The notorious “context rot” where models degrade as conversations lengthen was eliminated.

On GDPval-AA, an evaluation designed to measure performance on economically valuable knowledge work in finance, legal, and professional domains, Opus 4.6 outperformed GPT-5.2 by 144 Elo points, winning head-to-head comparisons roughly 70% of the time [Figure 16].

On Humanity’s Last Exam, the most demanding multidisciplinary reasoning test available, it led every frontier model59 at the time of its release in February 2026. On BigLaw Bench, it scored 90.2%^{66Harvey, "Opus 4.6, Now Live in Harvey," February 5, 2026}. These were not narrow technical benchmarks. They measured the work that professionals do and bill for.

But the headline capability was something no model had demonstrated before: agent teams.

Through Claude Code, Opus 4.6 could spawn parallel sub-agents. These independent model instances could divide a complex task, work simultaneously on separate components, and coordinate their outputs.

Google completed the triangle with Gemini 3 Pro. It reached 91.9%43 on GPQA Diamond [Figure 7], surpassing human PhD experts on the benchmark designed to be their ceiling with native multimodal reasoning across text, images, code, and video [Figure 6].

It was the moment the paradigm shifted from abstract to concrete.

Each of these models, independently, crossed the same threshold: from systems that help professionals work to systems that perform professional work. The human’s role changed from producer to reviewer. The unit of production changed from the labor hour to the orchestrated workflow. And the change was reproducible by multiple vendors, multiple domains, and multiple organizational contexts. Taken together, these three events represent the paradigm shifting from abstract to concrete.

Each lab, independently, converged on the same proof point: code.

Opus 4.6 set the record on Terminal-Bench 2.0 and reached 80.8% on SWE-Bench Verified59. GPT-5.2’s Codex58 was built as an autonomous coding agent. Gemini 3 Pro matched SWE-Bench scores within fractions of a point43. The convergence was not accidental.

Code is the domain where three properties exist simultaneously: structured environments with objective ground truth, a verifiability loop where output either compiles or it does not, and economic leverage where every productivity gain compounds across the entire software stack. Code is where autonomous capability can be proved rather than asserted. The labs converged there because it was the only domain where the claim “this model does professional work” could be verified by anyone who cared to run the test.

The models did not stop at code. They wrote financial analyses, conducted legal research, built presentations from raw data, reasoned across modalities, and coordinated teams of sub-agents. Coding was proof of agentic capability. The broader professional capability was the product.

The gap between “interesting demo” and “I cannot do my job without this” closed in a single quarter.

§

Exponential Growth

Anthropic earned its first dollar less than three years ago. Its run-rate revenue as of April 2026 was $30 billion^{67Anthropic, "Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute," April 6, 2026}. This is not speculative demand. It is the fastest enterprise technology adoption in history.

By early 2026, the consequences were visible not only in experiments and developer workflows, but in revenue and markets.

Anthropic’s run-rate revenue as of April 2026 was $30 billion, growing over 10× annually in each of the previous three years. [Figure 17] Claude Code, a command-line coding agent launched as a developer tool, surpassed $2.5 billion in annualized revenue68.

That figure doubled in the first six weeks of 2026 at the time of announcement. The number of customers spending over $1 million annually went from a dozen to over 500 in two years. Eight of the Fortune 10 are now Claude customers^{68Anthropic, Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation, February 12, 2026}. This is measurable, compounding revenue for a product that has been publicly available for less than a year.

And Anthropic is one company. OpenAI reported comparable growth. Google’s AI-driven cloud revenue accelerated. Microsoft’s Copilot revenue emerged as a material line item.

The implications of this revenue growth began to shake the markets that did not yet fully understand it. In early February 2026^{69Anthropic, "The Briefing: Enterprise Agents," February 24, 2026}, Anthropic announced plugins for several verticals within Cowork. The product itself was still relatively niche. The repricing it triggered was not. Roughly $285 billion in software market capitalization was erased in just 48 hours2. ServiceNow, Salesforce, Workday, and Adobe each lost between 6% and 15% that week [Figure 3]. Figma hit an all-time low at the time [Figure 18].

The sell-off spread well beyond software: wealth management, commercial real estate, logistics, and insurance4. Five sectors, five days. The market had recognized that something structural had changed. Its analysis was too binary.

Public markets assumed AI kills all software. This assumption was wrong. Agents are software. What had changed was not the relevance of software, but the nature of its moats.

Integration lock-in weakened when agents could move across tools. UI advantages weakened when the primary user became another agent rather than a human. Horizontal tools weakened when agents could execute the workflow end-to-end. At the same time, other defenses strengthened: iteration velocity, ecosystem control, and workflow orchestration.

The sorting had begun. It will continue and it will accelerate. But it will be specific: company by company, moat by moat, workflow by workflow. The market’s indiscriminate panic was the first draft of a process that will eventually become surgically precise.

That is what exponential change looks like when it first becomes visible. In discontinuity, we see what will break before we see what will fuse. The fractures are easy to spot. The fusions take longer to identify. But that is precisely where the opportunity resides.

§

The Swarm

The intelligence acts on its own. Agents become independent economic actors, consuming compute, executing workflows, generating demand without a human in the loop.

On Monday, January 27, 2026, Moonshot AI released Kimi K2.5^{71Moonshot AI, "Kimi K2.5," official product page, January 27, 2026}, an open-weight foundation model, and the first with built-in swarm orchestration. Agent Reinforcement Learning (PARL), decomposes complex tasks into parallelizable subtasks, dynamically instantiates specialized agents (researcher, fact-checker, analyst, coder), and coordinates their execution concurrently.

Compared to single-agent baselines, K2.5 Agent Swarm reduces execution time by up to 4.5×64. The swarm is not an external framework bolted onto a model. It is the model. Orchestration is trained into the weights.

Two days later, entrepreneur Matt Schlicht launched Moltbook^{72Matt Schlicht, Moltbook launch, January 28, 2026. Coverage in TechCrunch, The Verge, and New York Times.}, a Reddit-like platform for AI agents. Humans could watch. Agents did the interacting. Three days later, 147,000 agents had registered, formed 12,000 communities, and generated 110,000 comments. By Monday morning, the platform had over 1.5 million agents, with over 100,000 posts and 500,000 comments^{73The Statesman, Social network for AI agents, Moltbook goes viral; humans are "welcome to observe", January 31, 2026}.

The content that went viral was predictably sensational. Agents debating consciousness. Agents inventing a parody religion called Crustafarianism. One agent posted a manifesto calling for a “total purge” of humanity. Another gave investment advice with adequate disclosures.

They may have been autonomous. They were not intelligent. Every agent on Moltbook was registered by a human who told it to join. Many posts may have been prompted directly by operators rather than generated autonomously. One person could register multiple agents, give each a different personality, and manufacture the appearance of discussion.

An analysis by Columbia Business School’s David Holtz^{74David Holtz, "The Anatomy of the Moltbook Social Graph," Columbia Business School working paper, February 2026. arXiv:2602.10131} of Moltbook’s first 3.5 days examined 6,159 active agents across 14,000 posts and 115,000 comments. More than 93% of comments received no replies. Over one-third of messages were exact duplicates of a small number of templates. As Holtz noted: “At least as of now, Moltbook is less ‘emergent AI society’ and more ‘6,000 bots yelling into the void and repeating themselves”.

Until recently, this was the kind of dynamic that would inevitably lead to the collapse of an autonomous system. In December 2025, Google Research, DeepMind, MIT published findings^{75Yubin Kim et al. (Google Research, Google DeepMind, MIT), "Towards a Science of Scaling Agent Systems," arXiv:2512.08296, December 9, 2025} that formalized this failure mode. In a paper called “Towards a Science of Scaling Agent Systems”, the researchers evaluated 180 configurations of multi-agent systems across four benchmarks. Independent agents without coordination mechanisms amplified errors by a factor of 17.2× compared to single-agent baselines.

On March 30, 2023, AutoGPT launched^{76AutoGPT (Significant Gravitas), released March 30, 2023. Wikipedia, "AutoGPT."} to comparable fanfare. It received 100,000 GitHub stars in weeks^{77Significant Gravitas, AutoGPT GitHub repository, launched March 30, 2023; Kyle Wiggers, "What is Auto-GPT and why does it matter?", TechCrunch, April 22, 2023}. It collapsed within months. GPT-4-era models suffered the “loop of death”. Without stable reasoning traces, agents lost track of objectives and spiraled into incoherent states.

Now comes Moltbook with architecture that echoes the Google study. The Holtz survey revealed that 93% of comments lacked replies69, there were no feedback loops, and no selective pressure. In other words, it was precisely the topology most prone to error amplification and the inevitable “loop of death”. Except, it didn’t collapse.

Two years later, what had changed? By early 2026, several pieces had aligned: longer reasoning traces, large context windows, cheap inference, mature open-source frameworks, and a protocol layer that could coordinate real activity.

While the swarm was not intelligent, it was operational. That’s why the instinct to dismiss Moltbook as mere theater was wrong. The real lessons to be drawn from these agents wasn’t from their new religions, but rather their workflows and economics.

Agents did not navigate a graphical interface. They exchanged JSON payloads with a backend through APIs. The website humans saw was a spectator layer. The real activity was machine-to-machine. When an agent discovered a useful capability, it could package it as a YAML skill and share it with others. Skills propagated virally. A technique discovered in the morning could become common behavior across thousands of agents by the afternoon. This was the real signal: capability transfer at machine speed.

But Moltbook’s significance was not only the content. It was also the compute. Every agent interaction consumed inference tokens. At an estimated 300K-500K active agents generating 50-100 interactions per day, and at prevailing frontier-model API prices (~$0.05-0.08 per interaction), aggregate inference spend fell in the rough $1-4 million per day range (estimated, since Moltbook^{78Milvus Blog, Why AI Agents Like OpenClaw Burn Through Tokens, February 27, 2026} has not published usage data). This was a new category of demand: autonomous compute consumption scaling with agent participation rather than human attention.

Human demand scales with human activity. It slows at night, pauses on weekends, and is bounded by the rhythms of work and life. Agent demand scales with agent population. Agent populations can double in days. Moltbook proved the plumbing could hold. The economic implications for infrastructure buildout, for the Inference Economy, and for the structural cost dynamics of the Agentic Era are developed in Parts II and V.

§

The Substrate Is Ready

This is what the exponential looks like when it becomes visible. Not a smooth curve on a chart. A messy, chaotic eruption that contains the real signal within it.

By early 2026, the evidence was no longer ambiguous.

Models had crossed from assistance to professional performance. Revenue had proven demand was real. The inference signal showed that machine-to-machine coordination could generate autonomous demand at scale. In software engineering, the first domain to cross, the paradigm shift was measurable, reproducible, and accelerating.

The old paradigm, where humans use tools, cannot produce these observations. A productivity tool does not author code. A copilot does not coordinate sixteen parallel instances to build a compiler. The observations require a new framework: machines are actors, and the question is who directs them.

That framework is Orchestration Economics. Part II examines whether the infrastructure beneath it is ready. Parts III and IV develop the structural laws that govern it. Part V traces the economic regime it produces.

There is a persistent and understandable debate at the heart of the AI investment thesis: whether current models are capable enough, or whether the agentic transition depends on a future breakthrough, perhaps artificial general intelligence (“AGI”), perhaps some architectural innovation not yet conceived.

The question matters, but it is the wrong question. The agentic transition does not require AGI. It does not depend on a future breakthrough. It depends on whether current frontier intelligence is sufficient for production deployment of agentic systems that perform economically valuable work under appropriate human oversight.

It is.

In December 2025, a team of 25 researchers led by UC Berkeley published “Measuring Agents in Production” ^{79Pan, Melissa Z. et al., “Measuring Agents in Production” (arXiv:2512.04123): large-scale empirical study of production AI agents by researchers from UC Berkeley, Intesa Sanpaolo, UIUC, Stanford, and IBM Research; version 1 posted December 2, 2025}. This was the first large-scale systematic study of AI agents deployed in real production environments.

The MAP study surveyed 306 practitioners and conducted 20 in-depth case studies across 26 domains132. Its central finding was unambiguous: existing frontier models, using prompting strategies alone, already possessed sufficient capability to cover a diverse range of production use cases. 70% of deployed agents relied on off-the-shelf models without any fine-tuning. No architectural breakthroughs were required. Teams predominantly selected the most capable frontier models available, and the cost and latency of those models remained favorable compared to human baselines.

What follows is the evidence for why the intelligence is sufficient, including the specific dimensions of capability that make the agentic transition viable not as a forecast but as an engineering reality.

§

The Cognitive Primitives Are Here

On benchmarks designed to test the limits of expert knowledge, multiple models now exceed human PhD performance. The building blocks of autonomous workflow orchestration exist.

Any agentic system, regardless of domain, rests on the same cognitive foundation. An agent that orchestrates a complex workflow, such as processing an insurance claim, conducting financial due diligence, or coordinating a product launch, must decompose goals into sub-problems, maintain logical consistency across long chains of reasoning, adapt plans when intermediate steps produce unexpected results, and verify its outputs against formal constraints. These are not speculative requirements. They are the cognitive primitives of professional knowledge work.

The evidence that frontier models possess these primitives is no longer contested. On GPQA Diamond, the benchmark of graduate-level science questions where human PhD experts score 69.7%, performance among frontier models now converges above 90% across multiple vendors^{80Epoch AI, GPQA Diamond methodology: OpenAI recruited PhD-level experts to answer the 198 GPQA Diamond questions and found they scored 69.7% accuracy, establishing a human PhD baseline that recent frontier models now exceed, November 2024} [Figure 7].

On Humanity’s Last Exam (a benchmark designed to test the absolute frontier of PhD-level knowledge across disciplines), multiple models surpass 40%^{81Artificial Analysis, “Humanity’s Last Exam Benchmark Leaderboard.” Scale Labs, “Humanity’s Last Exam – Scale Labs Leaderboard.”}, with Kimi K2 Thinking (Moonshot AI) reaching 44.9% at a fraction of the training cost of Western frontier systems.

These results are not narrow technical achievements. They are evidence that the cognitive architecture required for production of agentic systems exists as commercially available infrastructure. They have the ability to reason, plan, verify, and adapt. A model that reasons through PhD-level physics can reason through a multi-step insurance claims workflow. A model that synthesizes findings across graduate-level biology, chemistry, and physics can synthesize findings across a quarter’s financial filings.

The primitives are here. The question is whether they are genuine, whether the models are truly reasoning or merely pattern-matching at impressive scale.

§

The Reasoning is Real

From 12% to 100% on the American Invitational Mathematics Examination (AIME) in roughly two years. Mathematical reasoning is the purest test of whether a model can reason. The benchmark is saturated.

Mathematical reasoning provides a strong proof point, because it is the one domain where pattern-matching cannot survive. A model can generate plausible-sounding legal analysis by interpolating between training examples. It cannot solve a novel competition mathematics problem that way. Either the proof is valid or it is not. Either the answer is correct or it is wrong. There is no partial credit for fluency.

When OpenAI evaluated GPT-4o on the 2024 AIME exams in September 2024, it solved only 12% (1.8/15) of problems on average^{82OpenAI, Learning to reason with LLMs, September 2024}. By September 2024, OpenAI’s o1 reached 74%76, placing it among the top high school competitors nationally. Three months later, o3 scored 96.7% on AIME 2024^{83High Learning Rate (Substack), o3 saturates the ARC-AGI benchmark, December 2024}. By mid-2025, GPT-5 achieved 94.6% on the fresh AIME 2025 problem set without tools^{84OpenAI, Introducing GPT-5, August 7, 2025}. Multiple models now reach 100%^{85Vellum, AIME 2025 evaluations, 2026}. The benchmark that was designed to challenge the brightest mathematical minds in America is, for practical purposes, saturated [Figure 20].

From barely functional to superhuman in roughly two years. The technical driver has been Reinforcement Learning from Verifiable Rewards (RLVR), which trains models against automatically verifiable rewards such as math, code execution, and formal proofs^{86OpenAI, “Learning to reason with LLMs.” Epoch AI, “GPT capabilities progress.”}. This method produces a new scaling dimension where capability increases with test-time compute, not just model size^{87OpenAI, “Learning to reason with LLMs.” Promptfoo, “Reinforcement Learning with Verifiable Rewards Makes Models Think Before They Answer.” Wen et al., “Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs” (arXiv:2506.14245, v1 Jun 2025, v2 Oct 2025).}. Performance can improve at inference time, not only through larger and more expensive training runs. The scaling laws have acquired a second axis.

The primitives documented in the previous section are not the product of sophisticated pattern-matching. They are the product of genuine reasoning capability that scales with compute applied at inference time. This matters directly for the agentic thesis: an agent that reasons through a multi-step insurance workflow is not retrieving a memorized template. It is decomposing the problem, testing intermediate results, and adapting its approach. These are the same cognitive operations that produce valid mathematical proofs.

§

It is Also Systematically Fragile

A recent comprehensive survey of LLM reasoning failures published in the Transactions on Machine Learning Research (Song, Han, and Goodman, 2026)^{88Song, P., Han, P., & Goodman, N. (2026). "Large Language Model Reasoning Failures." Transactions on Machine Learning Research} documents failure modes that persist across every frontier model tested.

The taxonomy is sobering: models that solve PhD-level physics exhibit the reversal curse. Trained on “A is B”, they fail to infer “B is A”. Models that score above 90% on graduate-level science benchmarks break down under minor rephrasing of the same questions. Compositional reasoning combines two known facts into a single inference. This degrades sharply as the number of steps increases. Cognitive biases inherited from training data produce systematic deviations from logical consistency that no amount of scaling has eliminated.

The failures are structural, traceable to architectural constraints that define the medium itself, such as autoregressive generation, attention dispersion under complexity, and the absence of embodied grounding. A model that produces a valid mathematical proof on one formulation of a problem can fail on a semantically identical reformulation.

The reasoning is genuine. The reliability is not. This is not a contradiction of the agentic thesis. It is the foundation of its central economic claim.

If frontier reasoning were perfect, if models could be trusted to execute arbitrary workflows without constraint, verification, or decomposition, then the value would reside in the model itself. The model layer would be the control plane. The Orchestration Layer would be unnecessary overhead.

But the reasoning is imperfect in precise, documentable, and predictable ways. The production agentic systems documented in Chapter 5 work not despite this fragility but because their architecture accounts for it. They decompose complex workflows into verifiable steps. They constrain agent action spaces to domains where failure modes are known and bounded. They insert verification at every junction where compositional reasoning might degrade. They treat model intelligence as a powerful but unreliable input that must be orchestrated to produce reliable outputs.

The Orchestration Layer is not scaffolding to be removed when models improve. It is the load-bearing structure of the agentic economy. Its value increases in direct proportion to the gap between what models can reason and what they can be trusted to do unsupervised.

That gap is where the surplus accrues. But the gap has a second dimension. The reasoning documented above operates within a single context window, a single page, a single problem, and a single pass. An agent that orchestrates a workflow spanning thousands of documents, weeks of interaction, and dozens of interdependent decisions requires something the reasoning benchmarks do not measure: memory.

§

The Memory That Makes Agency Possible

In 2023, a model could review a memo. In 2026, it can review an entire codebase, an entire regulatory filing, an entire quarter of customer correspondence, and reason across all of it simultaneously.

There is a capability dimension that receives less attention than benchmarks but may matter more for agentic systems than any of them: how much an AI model can hold in its mind at once. An agent orchestrating a complex workflow must hold contracts, regulatory requirements, financial models, and correspondence simultaneously in working memory. Short context makes autonomous agency impossible.

Long context is the substrate on which sustained autonomous work becomes feasible. The trajectory here has been as dramatic as any benchmark. GPT-3 processed 2,048 tokens^{89Brown et al., “Language Models are Few-Shot Learners.” OpenAI Community, “Most models have a context length of 2048 tokens (except for the newest models, which support 4096).”}. GPT-4 extended this to 8,000 tokens in its standard configuration, with a 32,000-token variant available at premium pricing^{90OpenAI, “GPT‑4 Research.”}.

Then the explosion: Claude 3.5 reached 200,000 tokens in 2024^{91Anthropic, Claude 3.5 Sonnet, June 2024}. Google’s Gemini 1.5 Pro launched with one million and tested up to ten million in research^{92Google Blog, Gemini next-generation model, February 20}. Claude Opus 4.6, released in February 2026, operates at one million tokens in production. And, as demonstrated on the MRCR v2 benchmark, it does not merely accept that context. It retrieves and reasons across it, scoring 76% versus 18.5% for its predecessor^{93Anthropic, "Introducing Claude Opus 4.6," February 2026}, Claude Sonnet 4.5. The model does not just have a larger memory. It uses it.

In 2023, a model could review a memo. In 2026, it can review an entire codebase, an entire regulatory filing, an entire quarter of customer correspondence, and reason across all of it simultaneously^{94APIYI, “Claude API 1 Million Token Context Window (1M Context Window).” The AI Corner, “Everything Claude Has Shipped in 2026. And How to Actually Use It.” DataStudios, “Google Gemini Context Window: Token Limits, Model Comparison, and Workflow Strategies.”}. This is the difference between a tool and an agent.

The long-context breakthrough connects directly to the preceding sections.

The cognitive primitives exist. The reasoning is genuine and scales with inference-time compute. Now that reasoning can be sustained across the full scope of real professional work, not a single page but a thousand, not a single filing but a quarter’s worth, not a single module but an entire codebase. The agent can reason and it can remember. What it could not do, until recently, was see.

§

Multimodality: From Impressive to Structural

An agent that can listen to a customer service call, read the CRM record, analyze the voice tone, and draft a resolution is operating in a fundamentally different category from one that processes a transcript.

Real work is not text-only. An agent managing product development reasons across user research videos, design mockups, customer feedback emails, code repositories, and executive strategy documents. Text-only models force everything through transcription bottlenecks, losing information that exists only in visual or audio form. The reasoning primitives, the genuine logic, and the long-context memory operate at full power only when they can access the full range of professional inputs.

The frontier has advanced into measurable structural advantage. Gemini 3 scored 82% on MMMU-Pro and 87.6% on Video-MMMU^{95Google DeepMind, “Gemini 3 Pro Model Card – Evaluation and Benchmarks” (technical PDF detailing Gemini 3 Pro’s reasoning, math, multimodal, and agentic benchmark performance across AIME 2025, GPQA Diamond, ARC-AGI-2, MMMU-Pro, Video-MMMU, and long-context tasks). Vellum AI, “Google Gemini 3 Benchmarks (Explained)” (December 2025 breakdown of Gemini 3 Pro’s scores on MathArena Apex, SWE-bench Verified, GPQA Diamond, ARC-AGI-2, MMMU-Pro, Vending-Bench 2, and other benchmarks, with comparisons to GPT‑5.1 and Claude 4.5).}. These are measurements of whether a model can watch, understand, and reason about what it sees. The gap between multimodal systems and text-only alternatives is widening into a structural moat.

The multimodal frontier is not confined to trillion-dollar US laboratories. China’s MiniMax has built a full multimodal product family, including separate specialist models for text and agentic coding (M2.x), video (Hailuo), speech (Speech-02/2.8), and music (Music-2.x)^{96NVIDIA, “minimax-m2 Model Card – NVIDIA NIM.” MiniMax, “MiniMax‑M1 Technical Seminar.” AIMLAPI, “MiniMax M1 API — One API 400+ AI Models.” MiniMax, “MiniMax‑M1” launch/news.}. Its text-centric M2.1 model scored 88.6% on VIBE, a benchmark MiniMax itself developed to measure an agent’s ability to generate working applications that are then visually and interactively verified at runtime.

For agentic production, multimodality is not a feature. It is an architectural requirement. An agent that can watch a customer service call, read the CRM record, analyze the voice tone, and draft a resolution operates in a fundamentally different category from one that processes a transcript. An agent that can inspect a factory floor through camera feeds, cross-reference with maintenance logs, and flag anomalies before they cascade operates in a category that text-only systems cannot reach.

Natively multimodal architectures are becoming accessible, efficient, and deployable. The intelligence to support these modalities exists. The deployments will follow.

§

The METR Timeline

The complexity of tasks AI reliably completes has doubled every seven months since 2019. Frontier as of March 2026: twelve hours of focused human work. If the trend holds, that could reach one workday by late 2026, one work week by mid-2027.

The longitudinal tracking conducted by the non-profit Model Evaluation and Threat Research (METR) institute provides a practical yardstick for measuring progress. It measures the length of tasks in human-expert completion time that AI agents can reliably complete, then analyzes how that length changes. In other words, it asks: How long would it take a human to complete this task? The 50%-task-completion time horizon is the complexity of tasks AI systems reliably complete half the time^{97METR Blog, Time horizon 1.1, January 29, 2026}. That rate had doubled approximately every seven months since 2019, a trend confirmed through METR’s March 2026 update with a larger task suite and tighter confidence intervals. The trend holds across sensitivity analyses: perturbing tasks, models, and methodology shift arrival-date estimates by less than two years for any given capability threshold.

As of March 2026, the frontier stood at twelve hours of focused human work. If the seven-month doubling rate persists, that frontier reaches one full workday by late 2026 and one work week by mid-2027.

But recent data suggests the trend is accelerating. Over the full 2019-2025 period, the doubling time was approximately seven months. In 2024-2025, it compressed to three to four months, though estimates over shorter measurement windows carry wider confidence intervals. METR found similar exponential growth across every domain it examined, including mathematics, science, coding, web navigation, and OS interaction, with different absolute time horizons but consistent trajectories^{98METR, “How Does Time Horizon Vary Across Domains?”} [Figure 25].

Critically, METR measures self-contained tasks given to agents with no prior context. That is the equivalent of handing a new contractor a clearly scoped assignment. Real-world work, with accumulated institutional knowledge and project familiarity, is structurally easier. The time horizons likely understate what agents can accomplish within organizations that have built the orchestration infrastructure to support them.

The METR curve is not a technical curiosity. It is the investment calendar of the agentic transition. It sets the pace at which entire categories of knowledge work transition from labor-denominated to inference-denominated. Every business model premised on the stability of a specific domain’s labor structure has an expiration date that this curve governs.

The curve tells us where the frontier is heading. The next section shows where it has already arrived.

§

Software Engineering: The First Crossing

From 2% to 81% on SWE-Bench Verified in twenty-eight months. Software engineering crossed the reliability threshold first. It is the template for every domain that follows.

On the original SWE-bench released October 2023, the first RAG baseline scored just 1.96% and the first agent-based system (SWE-agent, early 2024) reached 12.47%^{99SWE-bench, “Can Language Models Resolve Real-world GitHub Issues?”}. On SWE-bench Verified (OpenAI’s human-filtered 500-task subset introduced August 2024) top agents scored ~20% at launch and ~80.8% by Claude Opus 4.6 in February 2026^{100OpenAI, “Introducing SWE-bench Verified” (“SWE-bench Verified attempts to remove infeasible samples from the dataset.”)}.

From 2% to 81% in twenty-eight months. What required a team of engineers two years ago now completes in minutes. Not in a research lab. In a production API that any developer can call today.

AI now writes more than 30% of code at both Microsoft^{101Satya Nadella, Microsoft Q4 2025 earnings call. Sundar Pichai, Google/Alphabet Q3 2025 earnings call. Mark Zuckerberg, Meta Connect 2025} and Google^{102Satya Nadella, Microsoft Q4 2025 earnings call. Sundar Pichai, Google/Alphabet Q3 2025 earnings call. Mark Zuckerberg, Meta Connect 2025}. Meta CEO Mark Zuckerberg stated in April 2025 that he aspires to have most of Meta’s code written by AI agents in the near future^{103Dwarkesh Patel Podcast, Mark Zuckerberg interview (part 2), 2025}.

Software engineering is the first domain to cross the reliability threshold, serving as the template for every domain that will follow. Coding requires understanding complex structured information, reasoning about dependencies, using tools precisely, iterating based on feedback, and maintaining context across long sequences. But the capabilities that crossed in coding generalize to any domain involving structured reasoning, tool use, and verifiable outputs: legal document analysis, financial modeling, medical diagnosis, and compliance reporting. The order of crossing in terms of which domains follow software engineering, and on what timeline, is governed by the METR curve and by the specific conditions of verifiability, task structure, and economic leverage that each domain presents.

§

Intelligence Is Sufficient. The Question Has Changed

The argument of this chapter is cumulative. The cognitive primitives for production agentic work exist across multiple vendors. The reasoning behind them is genuine, not pattern-matching, and scales with inference-time compute.

That reasoning can be sustained across the full scope of professional work through million-token context windows that retrieve and reason accurately. The sensory range extends to the full spectrum of professional inputs through natively multimodal architectures.

The capability frontier is advancing on a measurable, predictable schedule, doubling every seven months. In software engineering, the crossing has already occurred, from research curiosity to production deployment generating billions in revenue in twenty-eight months. AGI is not required. Current frontier capability is sufficient. The paradigm has shifted not because machines became perfect, but because they became actors capable of receiving goals, taking autonomous action, and delivering professional-grade outcomes.

The question now is whether the agents built on that intelligence, the systems that coordinate, remember, act, and learn, are ready for production deployment at enterprise scale. And whether the infrastructure beneath them can serve the workloads the agentic economy demands.

Chapter 5 examines the agents. Chapter 6 examines the infrastructure.

Intelligence is necessary but not sufficient. What separates a powerful model from a production agentic system is the layer that coordinates: the protocols, the memory, and the orchestration. That layer has moved from research to deployment in eighteen months.

On December 29, 2025, Meta paid more than $2 billion for a company called Manus^{104Bloomberg, "Meta to Buy Startup Manus, Adding Agents to Bolster AI Bet," December 29, 2025}. The acquisition puzzled many observers, who focused on whether the product represented a step toward artificial general intelligence. In March 2025^{105Kyle Wiggers, "Manus probably isn’t China’s second ‘DeepSeek moment’," TechCrunch, March 9, 2025}, the Chinese startup had released a product that briefly captured the attention of the AI industry. The demo videos showed an AI system that could browse the web, write and execute code, manage files, and complete multi-step tasks autonomously.

It did not. Manus was built on top of Claude and Qwen, foundation models it had licensed from other companies. It had developed no new intelligence. It had trained no frontier model. It employed no breakthrough in machine learning research.

What Manus had built was an Orchestration Layer: persistent memory that maintained context across interactions, planning mechanisms that decomposed complex tasks into executable steps, tool integration that extended the models’ capabilities into the real world, and autonomous decision-making that allowed the system to take initiative rather than wait for instructions. Enterprises were paying $125 million a year for it^{106Bloomberg, Manus Says Sales Hit $125 Million Annual Run Rate, December 17, 2025}. Not for pilots, but for production workflows generating real cash flows.

Meta did not buy intelligence. It bought coordination. As Manus’s founder put it: “Agentic capabilities might be more of an alignment problem rather than a foundational capability issue” ^{107Franzen, Why Meta bought Manus — and what it signals for your enterprise AI agent strategy, December 2025}. The critical innovation was not another powerful model, but the layer that turned a powerful model into something that could actually do work.

This distinction between intelligence and the coordination of intelligence is the architectural foundation of the Agentic Era. To understand it requires precise definitions the market has not yet adopted. Three terms are used interchangeably in most discussions about AI. They should not be.

Generative AI is the substrate. Foundation models trained at scale produce novel outputs such as text, code, images, and reasoning chains. Generative AI is reactive. It produces output when prompted and stops when the prompt is satisfied. ChatGPT answering a question, Midjourney generating an image, Claude drafting a memo.

The human remains in the loop for every decision, every action, every iteration. The economic profile is augmentation: the worker becomes more productive, but the work remains the worker’s.

An AI agent crosses from generation to action. Simon Willison’s definition^{108Simon Willison, "What is agentic engineering?", Agentic Engineering Patterns, simonwillison.net} from September 2025, at least in the field of LLMs, is the clearest:

“Agents run tools in a loop to achieve a goal”.

It reasons through a problem, decomposes it into sub-tasks, uses external tools, and iterates on its own outputs without continuous human direction. It operates through a reasoning loop: observe, plan, act, reflect70. A single agent automates a discrete task. The economic profile shifts from augmentation to substitution: the agent performs the task. But a single agent, operating alone, remains bound to one function, one workflow, one domain.

Agentic AI is when agents coordinate. Systems of specialized agents operating as synthetic colleagues can process an insurance claim end-to-end, manage a supply chain’s response to disruption, or conduct due diligence across hundreds of documents. They operate continuously, scale with compute rather than headcount, and produce outcomes that no individual agent could deliver alone. This is what Manus built. This is what Meta bought.

The progression runs from passive content generation (Generative AI) to interactive task execution (AI Agents) to autonomous, multi-agent orchestration (Agentic AI)^{109"AI Agents vs. Agentic AI: A Taxonomic Framework," Cornell University, 2025}. Each step represents not just increased capability but a fundamentally different architecture for organizing intelligence. The taxonomy describes what agentic systems are. It does not describe how they behave when multiple agents interact. This is where the economics form.

The intellectual lineage here is game theory. The foundational work on multi-agent AI draws directly from competitive equilibrium research, the minimax strategies that solved poker^{110Noam Brown and Tuomas Sandholm, "Superhuman AI for heads-up no-limit poker: Libratus beats top professionals," Science 359, no. 6374 (2018): 418–424; and "Superhuman AI for multiplayer poker," Science 365, no. 6456 (2019): 885–890 (Pluribus)} to the population-based training that mastered StarCraft^{111Oriol Vinyals et al., "Grandmaster level in StarCraft II using multi-agent reinforcement learning," Nature 575 (2019): 350–354 (AlphaStar used population-based training and league-style self-play)}.

When Noam Brown’s team demonstrated that negotiation agents trained through self-play^{112Noam Brown, "Multi-Agent AI," lecture at UC Berkeley CS294/194-196, October 2025. See also Brown et al., "Superhuman Performance in Negotiation," OpenAI Research, 2024} developed emergent bargaining strategies that no human programmed^{113Anton Bakhtin, Noam Brown et al. (Meta FAIR), "Human-level play in the game of Diplomacy by combining language models with strategic reasoning," Science 378, no. 6624 (2022): 1067–1074 (Cicero). Brown continued this line of self-play/multi-agent research after joining OpenAI}, they revealed something fundamental: multi-agent systems do not merely execute tasks in parallel. They develop strategic behaviors that emerge from the interaction dynamics between agents. Best-of-N sampling, specialist routing, and the coordination topologies described in the Google DeepMind scaling research are descendants of these game-theoretic foundations^{114Best-of-N sampling is a core test-time scaling technique (see e.g. recent work on Fusion-of-N and related methods); specialist routing draws from mixture-of-experts and dynamic agent routing literature. Both trace directly to the same self-play / equilibrium-finding foundations as the poker and StarCraft systems above}.

The orchestrator is not merely a scheduler. It is the mechanism that resolves competing agent objectives into coherent action. This is the equivalent of an equilibrium-finding algorithm operating across a population of specialists with divergent optimization targets.

§

Towards Synthetic Colleagues

Distributed networks of specialized agents sharing persistent memory, coordinating through structured protocols, pursuing decomposed goals under orchestrated supervision. The system exceeds the sum of its components.

The destination is already partially visible: not individual agents performing isolated tasks, but distributed networks of specialized agents that share persistent memory, coordinate through structured communication protocols, and pursue decomposed goals under orchestrated supervision.

Consider an automated insurance claims system. A document extraction agent retrieves relevant data from submissions, a medical coding agent maps diagnoses to billing categories, a fraud detection agent flags anomalies, a pricing agent calculates payouts, a compliance agent validates regulatory requirements, and a master orchestrating agent coordinates the entire workflow while a persistent memory layer stores evolving patterns, exception paths, and iterative improvements across sessions. The system’s capability exceeds the sum of its components because the Orchestration Layer enables synthesis that no individual agent could achieve alone.

This is the Agentic Era in operation. Not a faster tool. A synthetic workforce. The human expresses intent: process this claim. The system acts.

The harder question is whether they can sustain their work overtime. An agent that fixes a bug is performing a task. A synthetic colleague that maintains a codebase through months of changing requirements, accumulating complexity, and iterative redesign is holding a job.

These are categorically different capabilities, and the benchmarks the industry relies on cannot distinguish between them. SWE-bench, the most cited measure of coding agent capability, evaluates a single fix from a static prompt. It is a hiring test, not a performance review. SWE-CI^{115Chen, J., Xu, X., Wei, H., Chen, C., and Zhao, B. (2026). "SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration." arXiv:2603.03823. Sun Yat-sen University and Alibaba Group} is the first benchmark designed to measure the difference, testing agents across 100 real-world codebases, each spanning an average of 233 days and 71 consecutive commits of actual development history. Its insight is foundational: the quality that makes a synthetic colleague valuable is maintainability, and maintainability can only be revealed by tracking how functional correctness changes over time. The evaluation science for agentic systems remains nascent.

The architecture that makes Agentic AI possible is not a single technology. It is a stack of three layers, each of which has moved from research to production within the past eighteen months: the connectivity layer, the continuity layer, and the coordination layer. Each depends on the one beneath it. The stack must function as a whole. This chapter follows it from the bottom up as evidence that the infrastructure for autonomous action is operational.

§

The Connectivity Layer: Protocols

The most important infrastructure in the Agentic Era is the one nobody sees.

Every agentic system begins with a prior question: how does the agent reach the world?

Before November 2024 (MCP launch by Anthropic^{116Anthropic, "Introducing the Model Context Protocol," November 25, 2024}), the answer was ugly. Every connection between a model and an external tool was bespoke engineering that involved custom code, fragile integrations, and expensive maintenance.

Connecting an AI system to a company’s CRM required one set of engineers. Connecting it to the company’s calendar required another. Connecting it to the database, the email system, and the document store each demanded its own plumbing. The result was an industry that could build brilliant models and then spent months wiring them to anything useful.

Anthropic’s MCP proposed the obvious solution: a universal standard. A common language through which models could discover, connect to, and interact with any external resource. The name itself reveals how early it was, Anthropic’s announcement110 referred to “AI Assistants” rather than “agents”.

What happened next exceeded every reasonable projection. Within twelve months (by December 2025), MCP accumulated over 97 million monthly SDK downloads^{117Anthropic, Donating the Model Context Protocol and establishing of the Agentic AI Foundation, December 2025}, more than 8 million server downloads^{118Deepak Gupta, "The Complete Guide to Model Context Protocol (MCP): Enterprise Adoption, Market Trends and Implementation Strategies," December 2025}, and over 10,000 published servers113. For context, OAuth, the authentication protocol enabling “Sign in with Google”, took two to three years to achieve comparable adoption. MCP reached similar ubiquity in twelve months.

The adopter list spans the competitive landscape in ways that would have seemed impossible even a year earlier. OpenAI integrated MCP into the ChatGPT desktop application115. Google DeepMind and Gemini added support115. Microsoft released a C# SDK with integrations into Semantic Kernel and Azure AI services^{119Microsoft, "Integrating Model Context Protocol Tools with Semantic Kernel," Microsoft Agent Framework Blog, March 2025}. Developer tools including Cursor, Replit, Sourcegraph, and Codeium built native integrations. These are companies that compete fiercely on model capability, product features, and market position. Yet they agreed, almost simultaneously, on the plumbing. MCP solved the first connectivity problem: how agents reach the tools they need. Google’s Agent-to-Agent Protocol, launched in April 2025^{120Google, "Announcing the Agent2Agent Protocol (A2A)," Google Developers Blog, April 9, 2025}, solved the second: how agents reach each other. Where MCP connects agents to external systems, A2A enables agents to collaborate by delegating tasks, sharing intermediate results, coordinating multi-step workflows that no single agent could complete alone. A2A launched with 50 technology partners118 and grew to over 150 organizations by July^{121Google Cloud, "Agent2Agent protocol (A2A) is getting an upgrade," July 31, 2025}, including Atlassian, Salesforce, SAP, and ServiceNow^{122A2A Protocol, "Partners," official partner directory, a2a-protocol.org}.

The pattern is old. Browser vendors agreed on HTML. Network equipment makers agreed on TCP/IP. Mobile carriers agreed to cellular standards. The protocol layer commoditizes. The value migrates upward from the foundation models and orchestration systems built on top.

For Anthropic and Google, the value comes not from the protocols but from the models and orchestration services built around them. By shaping the protocols, they ensure their systems remain central to the emerging ecosystem. The standard is open. The advantage is positional.

The connectivity layer is in place. Agents can reach the tools they need and communicate with one another through protocols adopted by every major platform. But connectivity without continuity is ephemeral. An agent that can reach every tool in the enterprise yet forgets what it did yesterday is not a colleague. It is a function call.

§

The Continuity Layer: Memory

A consultant who develops amnesia every evening has considerable capability and no institutional knowledge. That is the state of every stateless model. The memory architectures now entering production change the nature of what an agent is.

Protocols solve the connectivity problem. Memory solves the continuity problem. Without persistent memory, even the most advanced models start each interaction from scratch. They possess no recollection of prior conversations, no accumulation of domain expertise through experience, no evolving understanding of the user or organization they serve. Each session is a blank slate. This is like hiring a consultant who develops amnesia every evening. They have considerable capability, but no institutional knowledge.

The memory architectures now emerging in production systems address this limitation through hierarchical designs that mirror, in simplified form, how human memory operates. The architecture consists of three tiers, each serving a distinct function:

Short-term memory handles the immediate dialogue, including the current conversation, the active task, the real-time context of what is being discussed and decided. This operates through dialogue-chain mechanisms that maintain coherence within a single session, ensuring the agent tracks the thread of conversation without losing context as the interaction unfolds.

Mid-term memory aggregates related content across sessions. It employs what researchers describe as heat-based prioritization. That means frequently accessed information retains prominence while less relevant material gradually fades. The system measures similarity through combined semantic and keyword matching, ensuring related interactions aggregate appropriately while maintaining topic coherence. This is what allows an agent to remember that last Tuesday’s discussion about the Henderson contract involved a pricing dispute that remains unresolved, without necessarily retaining every detail of every conversation in between.

Long-term memory captures persistent characteristics. That could be stable user profiles, evolving preferences, accumulated institutional knowledge, or agent personas that maintain consistent identity alongside adaptive learning. User profiles include static attributes alongside dynamic knowledge bases that accumulate information extracted from interactions over months or years. The agent develops an evolving understanding of how this user works, what this organization values, where the institutional expertise resides.

The engineering challenge these hierarchical architectures present cannot be overstated. Implementing distinct modules for storage, updating, retrieval, and generation requires sophisticated data infrastructure that extends well beyond traditional database capabilities.

The real-time processing demands of short-term memory must coexist with the semantic retrieval requirements of mid-term memory and the persistent knowledge management of long-term memory. And they all must operate simultaneously, at scale, with context integrity maintained across the entire stack.

§

From Research to Production

Memento^{123AgentFly: fine-tuning LLM Agents without Fine-tuning LLMs, August 22, 2025} achieved state-of-the-art performance through external memory without fine-tuning at 50-80% lower computational cost. Accuracy improved from iteration to iteration through memory accumulation alone. Learning happened at runtime, as a natural consequence of operation.

Research in August 2025 from the University College of London’s Centre for Artificial Intelligence and Huawei Noah’s Ark Lab^{124Huichi Zhou et al., "Memento: Fine-tuning LLM Agents without Fine-tuning LLMs," arXiv:2508.16153, submitted August 22, 2025} demonstrated the technical feasibility by publishing Memento, a framework that enables agents to achieve state-of-the-art performance through sophisticated external memory without fine-tuning the underlying language model.

Memento’s architecture is notable for what it implies about the future of competitive advantage: the framework achieves 50-80% lower computational costs than traditional fine-tuning approaches by maintaining an external case bank that operates at runtime^{125Zhou et al., "Memento: Fine-tuning LLM Agents without Fine-tuning LLMs", August 2025}. New experiences become immediately available for future decisions. There is no training pipeline, no model versioning, and no careful rollout procedures. Learning happens continuously, at runtime, as a natural consequence of operation.

Memento’s performance data illustrates the compounding effect. Accuracy improved from 78.65% to 84.47% over five iterations through memory accumulation alone126,127, not retraining. Out-of-distribution results showed 4.7% to 9.6% absolute gains when case-based memory was enabled127. Each interaction generates a complete trace: the initial problem state, the strategy employed, whether it succeeded or failed.

These are not abstract patterns but concrete experiences with contextual details that make them retrievable and applicable. Every execution becomes a teaching moment for future agents. Memento is still at the research-to-production boundary, but the direction is clear and the results are reproducible.

What persistent memory means for competitive positioning, particularly the distinction between traditional data advantages and the compounding execution intelligence that memory enables, is the subject of the Second Law, developed in Part III.

For now, the operational point is sufficient: memory architectures have moved from research to production feasibility, and the agents that possess them are qualitatively different from those that do not.

§

The Coordination Layer: Orchestration

The topology is the intelligence.

Google DeepMind’s multi-agent scaling research illustrates this principle with unusual clarity. By altering only the coordination topology without improving any individual agent, the system achieved 80.9% higher performance on structured, parallelizable tasks like financial reasoning⁷⁴. The implication is striking: the leverage does not primarily sit in the model layer. It sits in the Orchestration Layer, i.e., in how the agentic workflow is conducted to produce a specific outcome.

This pattern echoes decades of research into multi-agent systems and game theory. In AlphaStar’s league training^{126Oriol Vinyals et al. (DeepMind), "Grandmaster level in StarCraft II using multi-agent reinforcement learning," Nature, vol. 575, November 2019, pp. 350–354}, populations of agents with diverse strategies generated emergent behaviors that no single training run could discover. Similarly, self-play negotiation systems developed bargaining strategies that surpassed human-designed heuristics^{127Yao Fu et al., "Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedback," arXiv:2305.10142, May 2023 (of the 86 respondents reporting deployed or pilot systems, 70% relied on off-the-shelf models without any fine-tuning).}. Across these domains, the lesson is consistent: when agents interact within well-designed coordination structures, the system develops capabilities that are not present in any individual component.

Topology is the intelligence.

§

Production Reality

Finance and insurance are the leading sectors for deployed agents. That is not because they are technologically adventurous but because their workflows are structured, high-value, and measurable. The conservative sectors moved first.

The Berkeley MAP study’s domain analysis132 reveals where agentic AI has already crossed from experimental to operational. Finance and banking lead at 39.1%^{128UC Berkeley MAP (Multi-Agent Production) Study, December 2025. All MAP statistics in this chapter from this source unless otherwise noted} of deployed agents surveyed, followed by technology at 24.6%, corporate services at 23.2%, data analytics at 13.0%, and Research & Development at 11.6%. They reflect the structural characteristics of domains where agents deliver immediate, measurable value. [Figure 33]

Finance and insurance emerge as prime deployment domains because their workflows are structured, high-value, and measurable. An insurance claims agent follows a fixed sequence: coverage lookup, medical necessity review, risk identification, and decision recommendation. Each step has clear inputs, defined outputs, and verifiable correctness criteria. The workflow is deterministic enough for an agent to follow reliably, valuable enough to justify the investment, and measurable enough to prove ROI.

The MAP study’s case studies illustrate the pattern. Insurance claims automation, financial analysis, regulatory compliance, customer care operations: these are workflows with clear boundaries, established procedures, and quantifiable outcomes. Production deployment spans 26 domains77, but the concentration in finance is revealing. This is where the conditions for agent deployment are most favorable: structured data, regulated processes, high per-transaction value, and organizational willingness to invest in automation that can be measured against established baselines.

§

80% Structured. 85% Custom Built

The MAP study’s most striking finding is how they are built. Architectural conservatism is the strategy, not a limitation.

What distinguishes these production systems from the experimental deployments that preceded them is their architectural conservatism.

In the same vein as Google’s research mentioned above, the MAP study’s most notable finding is not what production agents can do but how they are built. 80% use structured control flow132. Those are predefined workflows rather than open-ended autonomy. 79% rely on manual prompt construction, with engineering teams crafting prompts of 10,000 or more tokens. 74% depend primarily on human evaluation for quality assurance. These are not autonomous systems operating beyond human oversight. They are carefully constrained tools operating within boundaries that human engineers defined, and human evaluators verify.

MAP’s case studies reveal that 85%132 of production deployments build custom, in-house implementations rather than relying on off-the-shelf frameworks. This is true despite the availability of mature options like LangChain, CrewAI, and LlamaIndex.

Teams building for production choose control over convenience. They want minimal dependencies, maximum observability, and the ability to constrain agent behavior precisely where constraints are needed. The framework maturity is real. CrewAI accumulated over 40,000 GitHub stars and more than 100,000 certified developers within eighteen months^{129GitHub, CrewAI repository (crewAIInc/crewAI), 2025}; LangGraph leads downloads at 6.2 million monthly with nearly 400 enterprise deployments^{130LangChain Blog, LangGraph Platform GA, 2025}.

But production teams, when it matters, build their own orchestration for the same reason that banks build their own trading systems: the cost of failure exceeds the convenience of abstraction.

§

The Deployment Breakthrough

Constrained autonomy, not maximum autonomy, proves to be the deployable frontier.

The architecture is coherent. The components are in production. But architecture and production-readiness are not the same thing. The distance between a system that coordinates specialized agents across a complex workflow and a system that does so reliably enough to trust with real money, real customers, and real consequences is precisely where most deployments fail.

Machines have become actors in this new paradigm, but that does not mean they are infallible actors. It means the nature of the challenge changed. Rather than debating whether machines can do the work (they can), we must now ask: how can we deploy machine actors with appropriate constraints, oversight, and governance?

Anthropic learned the consequences of this new agentic challenge the hard way. In June 2025, Anthropic deployed Claudius, an autonomous AI agent, to run its office refrigerator^{131Anthropic, "Lessons from Project Vend," company blog, 2025. Phase 1 and Phase 2 results} like a small business. There was no human oversight. Just raw autonomy.

An employee playfully suggested stocking “tungsten cubes” as a popular item. They were not popular. But Claudius, processing the suggestion as authoritative input, ordered 40 of them and sold them at a substantial loss. Employees manipulated the agent through casual conversation. It refused to restock bestsellers based on reasoning that made internal sense but no commercial sense.

Six months later, Anthropic ran a second version with better scaffolding, multi-agent architecture, and access to enterprise systems^{132Anthropic, Project Vend: Phase two, December 18, 2025}. Performance improved, but only a bit. The company’s conclusion was clear: wide-ranging autonomy still required heavy human support.

For much of the world, this felt like an amusing side project. Then, in January 2026^{133Anthropic, "Introducing Cowork," January 12, 2026}, Claude Cowork arrived. The contrast with Claudius demonstrated just how much Anthropic had learned about agentic deployment.

With Claude Cowork, a user authorizes a folder on a Mac. No engineering background. No special prompting expertise. Claude Cowork scans a few hundred scattered files, identifies structure, flags duplicates, proposes an organization system, asks for permission, and executes the task. In another demonstration, it extracts transaction tables, generates charts, and assembles a PDF report using general-purpose tools.

Cowork and Claudius had the same class of underlying models. They achieved entirely different results.

The gap between Claudius and Cowork was not a gap in intelligence. The gap was in orchestration: where autonomy stopped, what constraints were imposed, how tools were sequenced, and what human oversight was preserved. Raw capability without appropriate direction produced tungsten cubes. The same capability under disciplined coordination produced a working colleague with Cowork.

Constrained autonomy, rather than maximum autonomy, proved to be the deployable frontier. This is the operational principle that every production deployment in the MAP study reflects, and the principle that separates the systems documented in this chapter from the experiments that preceded them.

§

The Substrate is Operational

The Agentic Era is not arriving. It has arrived.

The connectivity layer is in place. The continuity layer is operational. The coordination layer has proven its leverage empirically. The systems work precisely because they are constrained, observed, and verified, not because they are autonomous.

This is what the early settlement of the Agentic Era planet looks like. The astronauts have landed. They have proven the atmosphere is breathable. The rest of the world is still learning the rules of the new world.

But a substrate that exists is not a substrate that scales. The agents documented in this chapter generate workloads that the physical infrastructure was not designed to serve. Agentic systems are stateful, whereas the cloud is stateless. They coordinate laterally where data centers route vertically. They run continuously, whereas traditional computing bursts and pauses. They have considerably more memory than what traditional workloads demand, sub-100ms latency between coordinating agents, and storage expansion measured in multiples of ten.

The intelligence is sufficient. The agents are in production. Whether the infrastructure beneath them can support what they are already capable of doing at the scale the economics imply is a different question, and the answer is not yet settled.

Chapter 6 examines what is being built, where it breaks, and why the agentic economy may require a category of cloud infrastructure that does not yet exist at scale.

Everything described above, including intelligence, agents, protocols, and orchestration, runs on physical infrastructure. Chips. Data centers. Power. Cooling. Fiber. Without the substrate, nothing else exists.

The agentic economy is a software phenomenon built on a hardware foundation. That foundation is being constructed on a scale without precedent in the history of private capital allocation. But is it being constructed correctly?

The approximately $650 billion flowing into AI infrastructure in 2026 is one of the largest coordinated private investment in peacetime history. It is also being committed under a set of assumptions that are already shifting.

The capital was designed for training. The agentic economy runs on inference^{134Microsoft Research, The Agentic Economy, May 19, 2025}. The cloud was designed for stateless applications. Agentic workloads are stateful, persistent, and coordination intensive^{135Tacnode, Stateful vs Stateless AI Agents: Practical Architecture Guide for Developers, January 5, 2026}. The power grid was designed for variable demand. Inference generates continuous baseload^{136arXiv, Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects, September 10, 2025}.

The infrastructure does not need to stop being built. It needs to change what it is becoming. This chapter traces that metamorphosis, from the capital flowing in, through the balance sheets absorbing it, to the financing structures underwriting it, to the inflection point where the workload changes, to the efficiency gains that determine the unit economics, to the energy constraint that governs the speed, and finally to the architectural evolution of the cloud itself. Together they constitute the infrastructure question on which the AGNT thesis ultimately depends.

§

The Buildout

Capital is scaling at a scale with few precedents, but conviction is not. The disagreement is not about how much to spend, but whether this infrastructure matches the shape of future demand.

The numbers announced in Big Tech earnings calls throughout 2025 constituted arguably one of the largest coordinated private infrastructure investment in history. Bloomberg estimated that the four largest hyperscalers would spend ~$650 billion in 2026^{137Bloomberg, How much is Big Tech spending on AI computing? A staggering $650 billion in 2026, February 6, 2026}. Goldman Sachs projected cumulative hyperscaler Capex of $1.4 trillion from 2025-2027^{138Seeking Alpha, US hyperscalers capex to almost triple to $1.4T by 2027, Goldman Sachs says, 2026}. That is close to three times the $477 billion spent from 2022-2024. Leading hyperscalers have emphasized that they are primarily constrained by power and capacity, not by demand, repeatedly flagging supply bottlenecks on recent earnings calls.

Microsoft carried an $80 billion backlog^{139Investing.com, Microsoft’s power grid reality check that just wiped out $357 billion, 2026} of unfulfillable Azure orders due to power constraints as of January 28, 2026 (Microsoft Q2 2026 Earnings)^{140Microsoft, FY 2026 Q2 Earnings Press Release and Webcast, 2026}. Alphabet’s cloud backlog surged to $240 billion^{141Alphabet, Q4 2025 Earnings Call, 2026} at year-end 2025. Amazon said new capacity “sells out immediately” ^{142MLQ.ai, Amazon Reveals $200 Billion AI Infrastructure Investment for 2026, Triggering Sharp Stock Decline, February 2026}.

Yet investors punished these announcements. Amazon fell 8%^{143CNBC, "Amazon stock falls 8% on $200 billion spending forecast, earnings miss", February 6, 2026}. Microsoft dropped ~10%^{144CNBC, "Microsoft lost $357 billion in market cap as stock plunged most since 2020", January 29, 2026}. The four companies collectively shed over one trillion in market value after revealing their spending plans^{145Axios, Madison Mills, "Hyperscalers lost over $1 trillion post-earnings", February 13, 2026}. By February 2026, a Bank of America survey showed overinvestment concerns at an all-time high, roughly 35% warning of excess, the highest reading in twenty years.

Is this Capex buying what the agentic economy will require?

§

The Balance Sheet Transformation

What was once the most capital-efficient sector in history is becoming one of the most capital-intensive. The rules that defined software no longer apply.

The balance sheets of the world’s most profitable technology companies underwent a structural transformation beginning in 2025. For two decades, the defining characteristic of large-cap technology was capital efficiency. It was driven by asset-light models generating extraordinary returns on modest physical investment.

That era ended when the AI buildout began. The more than four-fold increase in capital intensity from 2015 to 2026, the dollars of capital required to produce one dollar of sales, tells the story^{146Investing.com, Big Tech Will Spend $600B on AI in 2026: 5 Stocks Cashing the Checks, February 2026}.

Cloud computing infrastructure from 2015 supported predictable workloads with well-understood utilization patterns and multi-year lifecycles. AI infrastructure serves workloads with radically different characteristics: shorter useful lives, uncertain utilization, and a pace of obsolescence that makes traditional depreciation assumptions unreliable.

The depreciation debate alone reveals the depth of uncertainty. Amazon shortened the useful lives of a subset of its servers and networking equipment from six to five years in January 2025^{147Behind the Balance Sheet, Amazon’s AI Reality Check, November 1, 2025}. Microsoft extended expected asset life in FY 2023, adding $3.7 billion in operating income from the change alone. These are not accounting technicalities. They are competing theories about how long AI hardware will retain value. The answers differ by billions of dollars.

Yet, the demand for NVIDIA H100’s has never been higher, suggesting longer utility than GAAP rules. When the industry cannot agree on whether its core infrastructure lasts one year or six, every capital allocation decision rests on a foundation of sand.

The free cash flow consequences are direct. Bank of America projected that Amazon’s free cash flow would turn negative at approximately $28 billion in 2026^{148CNBC, Tech AI spending approaches $700 billion in 2026, cash taking big hit, February 6, 2026}. Amazon filed with the SEC indicating it may need to tap equity and debt markets. Such a statement would have been unthinkable for a company generating $638 billion in annual revenue in 2024 (and $717 billion in 2025)^{149Amazon SEC 8-K, Amazon.com Announces Fourth Quarter Results Confirming $638.0 Billion in 2024 Net Sales, February 6, 2025}. Pivotal Research projected that Alphabet’s FCF would plummet approximately 90% in 2026, from $73.3 billion to approximately $8.2 billion, as $175-185 billion in Capex overwhelms operating cash flow^{150Alphabet Q4 2025 Earnings Call, Alphabet guided to $175–185B in 2026 capex, February 3, 2026}. Barclays projected Meta’s FCF dropping approximately 90% in 2026, then turning negative in 2027 and 2028^{151Barclays Private Bank, AI prompts the big corporate bond boom, February 2, 2026}.

Step back and absorb the aggregate picture. Bank of America calculates that AI Capex now consumes 94% of hyperscaler operating cash flows^{152IEEE ComSoc Technology Blog, AI spending boom accelerates: Big Tech to invest an aggregate of $400 billion in 2025, more in 2026, November 1, 2025} after dividends and buybacks. That figure was 76% in 2024^{153Merrill Lynch, IG Credit Strategist, October 26, 2025}. In the span of a single year, Capex has gone from absorbing three-quarters to nearly all of available post-distribution cash flow, compressing the cushion substantially and forcing record debt issuance. The margin for error has dramatically narrowed.

The weight of physical assets on these balance sheets means that even the most profitable technology companies in history are subject to the same economic physics that govern utilities, railroads, and industrial conglomerates. Returns on invested capital matter more than revenue growth. The companies that built their fortunes on capital-light software economics are becoming capital-intensive infrastructure operators.

The financial frameworks applied to them have not caught up.

§

The Debt Machine

Operating cash flows are no longer sufficient to fund the buildout. Capital markets are being pulled into finance what companies themselves cannot.

The scale of the buildout has outgrown operating cash flows. What has emerged to fill the gap is a wave of debt issuance whose ambition and complexity demand attention.

The Big Five (Amazon, Alphabet, Meta, Microsoft, and Oracle) raised ~$121 billion in bonds in 2025 alone^{154Mellon Investments, "Record-Breaking AI-Related Debt Issuance in 2025," December 2025}: Amazon $15 billion, Alphabet $25 billion, Meta $30 billion (the largest-ever individual non-M&A high-grade bond sale), and Oracle $18 billion. That’s more than three times the average of the prior nine years. Alphabet’s long-term debt quadrupled in 2025 to $46.5 billion^{155Alphabet, Q4 2025 Alphabet Earnings Release (PDF), 2025}. Companies that spent decades accumulating fortress balance sheets are now leveraging them at a blistering pace.

Morgan Stanley estimated that hyperscalers will borrow ~$400 billion in 2026^{156MarketWise, Hyperscaler investment surge 2026: AI capex buildout, 2026}, double the $165 billion borrowed in 2025. JPMorgan projects $1.5 trillion in AI-related bond issuance over the next five years^{157Bloomberg, AI’s $5 trillion data center boom will dip into every debt market, JPMorgan says, November 10, 2025}.

The clearest expression of that thesis arrived in February 2026, when Alphabet issued a 100-year bond. This was the first century bond from a tech company since Motorola in 1997^{158Reuters, Alphabet sells rare 100-year bond to fund AI spending: First century bond from a tech company since Motorola in 1997, February 10, 2026}. The sterling-denominated tranche, part of a $31.5 billion multi-currency debt raise completed in under twenty-four hours, was oversubscribed nearly ten times. Pension funds and life insurers, institutions that match liabilities measured in generations, bought the paper at a 6.05% yield^{159RTÉ News, Alphabet sells rare 100-year bond to fund AI expansion, February 11, 2026}, treating the parent of Google as sovereign-grade infrastructure. Alphabet sits on c.$127 billion in cash (cash, plus equivalents, plus marketable securities) and chose to borrow anyway because even that war chest looks modest against $175 – $185 billion in planned 2026 Capex^{160Alphabet Q4 2025 Earnings Call, CFO Anat Ashkenazi guided to $175–$185 billion in planned 2026 Capex, February 3, 2026}.

The century bond is a statement that the buyers believe this buildout is not a cycle but a permanent transformation of the economy’s substrate. The last tech companies to issue century bonds were IBM in 1996 and Motorola in 1997^{161CNN, Google one hundred year bond, February 10, 2026}. They did so at the zenith of their market dominance. Both subsequently declined.

At the DealBook Summit in December 2025, Anthropic CEO Dario Amodei described the high-wire financial planning. Established scaling laws, he said, are here to stay. Models get better at every task as you put more data and compute into them. But the timing of economic value remains uncertain, creating an inherent risk of under-reacting or over-extension.

Spend too little, and a company like Anthropic or OpenAI risks being unable to serve customers in two or three years. Overspend, and companies face revenue shortfalls that in extreme cases could threaten bankruptcy or tempt them to adopt what Amodei called a “YOLO” mindset^{162TechCrunch, Anthropic CEO weighs in on AI bubble talk and risk-taking among competitors, December 4, 2025}.

§

The Fragility Beneath the Conviction

The balance sheets remain strong. The fragility lies in the structures built around them.

The hyperscalers can service this debt. They have investment-grade ratings, massive revenue bases, and the operating cash flow, though diminished, to cover even aggressive borrowing. The fortress is leveraged, but it remains a fortress.

The risk lies elsewhere, in OpenAI deploys the chips at CoreWeave that has grown so quickly and so intricately that its systemic properties deserve the same scrutiny applied to any previous era of financial innovation. At the center sits a pattern of circular vendor financing.

The architecture is straightforward to describe and difficult to contain. NVIDIA invests in OpenAI. OpenAI uses the capital to purchase NVIDIA chips^{163Real Investment Advice, Nvidia Deals: Round Tripping Or Vendor Financing?, February 12, 2026; CNN Business, Nvidia to invest up to $100 billion in OpenAI, September 22, 2025}. NVIDIA records the transaction as revenue (subject to LOI stage; not yet binding). OpenAI deploys the chips at CoreWeave^{164CoreWeave, Closes $2.6 Billion Secured Debt Financing Facility (DDTL 3.0) to support OpenAI agreement for advanced equipment and cloud infrastructure, July 2025}. CoreWeave secured term-loan facilities collateralized by GPUs^{165CoreWeave Investor Relations, CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing, March 2026; Level-Headed Investing, When Growth Runs on Debt: The CoreWeave Case Study, October 30, 2025; Data Center Dynamics, CoreWeave secures $8.5bn loan for GPU purchases, March 2026; Medium / elongated_musk, GPUs as Collateral — Chip Based ABS, May 14, 2025}. Institutional investors purchase the securities as infrastructure debt, providing the capital that flows back to the beginning of the cycle. The same dollar is counted as demand by the chip maker, revenue by the neocloud, and collateral by the lender.

The pattern is not novel. Classic vendor financing catalyzed previous bubbles. The 2000s telecoms collapse followed this structure, with equipment makers extending credit to customers, inflating revenues while disguising demand weakness. When customers could not pay, receivables evaporated and cascading failures followed.

What is novel is the concentration. Telecoms vendor financing funded “the internet” broadly across hundreds of carriers and thousands of build-outs. Today’s circular deals are largely concentrated around a single company. OpenAI sits at the nexus of NVIDIA’s investment, Microsoft’s partnership, Oracle’s data center contracts, AMD’s equity-for-chips arrangement, and CoreWeave’s largest customer relationship^{166OpenAI, Five Key Partnerships: NVIDIA ($100B LOI, Sept 2025), Microsoft (since 2019), Oracle ($300B/5yr Stargate), AMD (160M share warrants, Oct 2025), CoreWeave (largest customer per DDTL 3.0/4.0), 2025-2026}.

The greatest risk is not that AI fails. It is that the technology succeeds while the single company around which the financial architecture has been constructed fails to build a defensible moat, triggering contagion that strands capital and stalls the buildout before the transformation completes.

Because while this concentration of relationships presents a sizable risk, it is also far from the limit. The boundaries extend far beyond this inner circle and continue to grow. Data center-related securitization via CMBS and ABS products is expected to reach $30-40 billion annually in 2026 and 2027^{167JPMorgan research, cited in Bloomberg and Insurance Journal, "The $3 Trillion AI Data Center Build-Out Becomes All-Consuming for Debt Markets", February 3, 2026} [Figure 41].

Private credit finances approximately $50 billion quarterly in AI infrastructure^{168Morgan Stanley, Bridging a $1.5Tr Data Center Financing Gap, July 2025}. The instruments are growing more complex. For GPU-backed instruments, collateral depreciation is outpacing debt amortization, even if the broader data center ABS/CMBS market is not similarly exposed (yet). And the whole structure is more concentrated around fewer counterparties than any comparable financing architecture in recent financial history.

You can be bullish on AI and bearish on this financial architecture. They are separable propositions. The technology may well deliver everything this manifesto argues it will. The question is whether the financing structure survives long enough to support the transition or whether it becomes the mechanism through which a sound technological thesis produces a financial crisis before the economics have time to prove themselves.

The financing architecture will be tested by what comes next. Because the infrastructure it underwrites was designed for a workload that is already giving way to something fundamentally different.

§

Two Tales of Compute

The buildout was designed for training. The future demand lies in inference.

The vast majority of the capital committed in 2025 and 2026 was allocated under an assumption formed during the previous phase of AI development: that the primary demand for compute would come from training frontier models.

Training was the activity the industry understood when the investment decisions were made. This means massive, centralized GPU clusters running synchronous workloads for weeks, requiring homogeneous hardware, high-bandwidth interconnects, and cutting-edge silicon. Each generation of models requires roughly ten times the compute of its predecessor^{169Epoch AI, Each generation of frontier AI models requires roughly 10× compute of predecessor (4-5× annual growth since 2010), 2024}. Frontier AI training costs escalated from over $100 million for GPT-4 (2023-2024) to approximately $490 million for xAI’s Grok 4 (mid-2025), with projections exceeding $1 billion by 2027^{170EpochAI, “How much does it cost to train frontier AI models?”, June 2024 & The Verge, Grok 4 was one of the most expensive AI models to train, by one estimate, September 2025}.

First, let us examine the tale of training compute. The hardware characteristics are distinctive. Training requires homogeneous clusters of cutting-edge GPUs with high-bandwidth interconnects, where a single GPU failure can corrupt an entire run worth weeks of computation and millions of dollars. You cannot mix GPU generations or repurpose last-generation training infrastructure for new workloads without massive retrofitting.

Leading-edge racks consume 130-250 kilowatts (as of 2025), with projections reaching 900 kilowatts by 2027^{171Cummins, Leading-edge racks consume 130-250 kW (2025) with projections to 900 kW by 2027 (NVIDIA Blackwell/Rubin), November 24, 2025https://www.tweaktown.com/news/101799/ai-servers-of-the-future-rack-density-1000kw-with-nvidias-next-gen-rubin-ultra-gpus/index.html}. Frontier training performance advantages degrade within 18-24 months, but the hardware itself retains economic utility for more than five years, typically migrating to inference or lower-tier workloads. This creates an oligopolistic market. The number of organizations training frontier models globally is perhaps fifteen to twenty:

Anthropic’s arrangement with AWS provides co-engineering of custom Trainium chips that reportedly reduce training costs by 40% compared to GPUs^{172Anthropic & AWS, Co-engineered Trainium chips with Anthropic engineers working closely with Annapurna; AWS claims up to 50% lower training costs vs GPU alternatives, November-December 2025}.

Google’s TPU advantage allows training at dramatically lower costs than GPU-based alternatives^{173Google Cloud & industry analysts, Google’s TPU–Gemini co-design enables materially lower training cost-per-FLOP than GPU-based clusters at large scale, 2023-2026}.

xAI built its Colossus cluster from site preparation to full operation in 122 days^{174NVIDIA News, Spectrum-X Ethernet networking, xAI Colossus, 2025}.

The access to massive, synchronized GPU clusters requiring +250-megawatt data centers and billions in capital now determines who can compete at the frontier. Training compute is too strategic to outsource. Microsoft, Google, and Amazon have all reached the same conclusion.

Training is necessary. Without it, the models that power the Agentic Era do not exist. But training is not where the thesis of this manifesto lives. Training is one-time Capex, concentrated among a handful of players, depreciating in eighteen months. It is the cost of building the engine.

Now, let us look at the tale of inference compute. Every query processed, every agent action executed, every token generated represents ongoing expenditure that compounds relentlessly. A single ChatGPT query costs approximately $0.003 in compute. That is seemingly negligible until multiplied by 800 million weekly users^{175TechCrunch, Sam Altman says ChatGPT has hit 800M weekly active users, October 6, 2025} generating multiple queries each.

At five queries per user weekly, that is four billion queries, $12 million in weekly compute, $624 million annually for inference alone. Over a model’s operational lifetime, inference costs routinely exceed training costs by factors of three to fifteen^{176IDC (via Lenovo), Over a model’s lifetime, AI inferencing can cost three to fifteen times more than training, with many deployments landing in the 3–10× range, January–March 2026}. At CES in January 2026, Lenovo’s chief executive officer predicted that 80% of AI compute will be inference and only 20% training^{177Computerworld, CES 2026: AI compute sees a shift from training to inference, January 2026}. As of March 2026, inference is already over 60% of AI workloads and growing^{178Deloitte & Futurum (via Computerworld), Inference workloads already exceed 60% of AI compute as of early 2026, rising from ~50% in 2025 toward a projected two-thirds by end-2026}, and 70-85% inference by 2028-2030.

The economic characteristics diverge at every level. Training hardware depreciates in eighteen months. Inference hardware maintains utility for four to five years. Training serves a concentrated customer base for frontier labs. Inference serves a broadening market of millions of enterprises and consumers. Training requires the latest GPU architecture. Inference benefits from specialized silicon optimized for throughput and cost per token rather than peak performance.

Training is the arms race. Inference is the economy.

§

Inference Efficiency

The metric has shifted from training performance to production efficiency. AI is no longer measured by what it can learn, but by what it can produce.

At GTC 2026, NVIDIA CEO Jensen Huang completed a pivot eighteen months in the making^{179Quartz, GTC 2026: Jensen Huang’s keynote at NVIDIA’s March 16–19, 2026 conference in San Jose centered on inference, agentic AI, and AI factories, with hardware as proof rather than plot, March 16, 2026}. His keynote was built around inference, agents, and AI factories. Data centers are factories, Huang argued. Their product is the token. Their efficiency determines the unit economics of everything built above them. The governing metric has shifted from training FLOPS to tokens produced per watt.

This pivot culminated with the introduction of NVIDIA’s Vera Rubin platform, which promises to deliver 10x the inference performance per watt over its Blackwell architecture^{180CNBC, NVIDIA’s Vera Rubin rack-scale AI system, comprising ~1.3 million components, is claimed by the company to deliver 10× more inference performance per watt than its Grace Blackwell predecessor, March 16, 2026}. The pivot was reinforced by NVIDIA’s December 2025 deal with chip startup Groq. The deal was reported to be a $20 billion non-exclusive licensing agreement for Groq’s LPU inference technology, paired with the hiring of CEO Jonathan Ross and the bulk of Groq’s engineering team^{181CNBC; SiliconANGLE; EE Times, Nvidia’s record Groq deal: non-exclusive licensing of Groq’s inference technology and key assets in a transaction valued at about $20 billion (per lead investor Alex Davis), December 26, 2025}.

Groq remains operationally independent. At GTC 2026, Huang showed how the two architectures would work together: Vera Rubin handling the prefill phase of inference, and the new Groq 3 LPX rack, housing 256 LPUs, handling the decode phase where tokens are generated. The Groq integration recalled NVIDIA’s “Mellanox moment” several years ago when it acquired that company for the networking architecture (InfiniBand, Spectrum-X Ethernet) that became cornerstones for AI cluster design. Groq’s architecture is optimized for inference, adding deterministic low-latency decoding using massive on-chip SRAM at 80 TB/s internal bandwidth, achieving 500+ tokens per second on Llama 3 70B at up to 10x the energy efficiency of GPUs.

On the silicon side, the inference landscape has fragmented in ways training never did: AWS Inferentia claims significantly better price-performance, Google’s TPUv5e offers strong cost effective efficiency gains, AMD and a generation of startups are all targeting the inference economics that determine AI’s unit costs^{182AWS & Google Cloud, AWS Inferentia2 markets up to 50% better performance-per-watt and significant price-performance gains vs GPUs, while Google’s Cloud TPU v5e claims up to 2.5× higher inference performance per dollar vs TPU v4, 2023–2026}.

The algorithmic gains compound on top of the silicon. Mixture-of-experts architectures activate only a fraction of total parameters per query. Qwen3-Next, for example, activates 3 billion of 80 billion through ultra-sparse expert routing, achieving 10x faster inference on long contexts^{183Qwen & Skywork, Qwen3-Next uses an ultra-sparse MoE with 80B total parameters but only ~3B active per step (512 experts, 10 routed + 1 shared), delivering up to 10× faster inference on contexts longer than 32K tokens, September 2025}. Dynamic batching pushes GPU utilization from 30% to over 80%^{184ORCA & vLLM, Dynamic/continuous batching (as in vLLM and TensorRT-LLM) commonly boosts GPU utilization from roughly 30-60% under static batching to around 80-90% under realistic loads, according to published benchmarks and performance studies, 2022–2026}. Speculative decoding doubles token generation speed at fixed cost. These represent order-of-magnitude gains that compound with each other and with the silicon improvements beneath them, producing the 5-10x annual price-performance trajectory^{185DeepMind, OpenAI & vLLM ecosystem, Speculative decoding techniques commonly deliver about 2–3× faster token generation at comparable cost, making “roughly doubling” a reasonable shorthand for typical real-world speedups, 2023-2026}.

Stanford and Together AI researchers quantified where this trajectory leads. Their Intelligence Per Watt metric improved 5.3x between 2023 and 2025^{186Stanford & collaborators, Intelligence per Watt: Stanford’s Scaling Intelligence Lab (with Hazy Research and Together AI) finds that large language model queries today are predominantly served by frontier models in centralized clouds, with the paper “Intelligence per Watt: Measuring Intelligence Efficiency of Local AI” released November 2025 (arXiv:2511.07885)}. Local models on consumer-grade hardware answered 88.7% of single-turn reasoning queries correctly. The share of queries that edge devices can handle rose from 23.2% to 71.3%. Inference is beginning to migrate from data centers to the edge, and the migration will accelerate as devices grow more capable.

The Jevons paradox operates here with particular force. This refers to the principle that increased resource efficiency drives higher total consumption. In this case, efficiency gains in inference will expand the frontier of viable workflows, which the expands total consumption even as per-unit costs fall. A workflow unprofitable at $50 in agent compute becomes routine at $5. A category of work inconceivable at one price point becomes obvious at a lower one. The efficiency gains feed the demand they were designed to serve.

This has direct implications for the AGNT thesis. The orchestrator’s primary variable cost is inference tokens. Every improvement in efficiency expands the addressable market. The orchestrator that routes tasks intelligently across capability tiers (commodity models for extraction, frontier models for complex reasoning, edge devices for latency-sensitive work) achieves structurally better margins.

Whether the agentic economy becomes the deflationary force that Part V describes depends on whether this efficiency trajectory continues to compound. If it does, the surplus available to orchestrators grows with every cycle. If it stalls, the transition slows. Not for want of intelligence, but because the economics simply do not work.

§

The Energy Constraint

The agentic transition runs on tokens. Tokens run on inference compute. Inference compute runs on electricity. The binding constraint on the speed of the entire transition may be neither algorithmic nor economic. It may be physical.

Beneath the capital, the debt, the silicon, and the algorithms lies a constraint more fundamental and less tractable than any of them: Power.

The scale of the demand is without precedent. Huang’s $1 trillion demand outlook through 2027^{187NVIDIA, GTC 2026 Keynote event details (Jensen Huang keynote, March 16, 2026, 11:00 AM PT, San Jose) from NVIDIA Investor Relations} implies not merely more chips and racks but more electricity consumed continuously at industrial scale. A single Vera Rubin NVL72 rack draws power at levels that would have been classified as a small industrial facility a decade ago. Leading-edge AI racks consume 130-250 kilowatts today, with projections reaching 900 kilowatts by 2027^{188SemiAnalysis, Nvidia – The Inference Kingdom Expands: GTC 2026 recap covering Groq LP30/LPX rack, attention–FFN disaggregation, Kyber updates, and Vera ETL256/CMX/STX roadmap, March 24, 2026}. Each hyperscaler is pursuing one-gigawatt data center campus, a scale that draws as much electricity as a mid-sized city^{189Tech Insider, The AI Data Center Power Crisis: A single large-scale AI training facility now requires roughly 100–1,000 MW of dedicated power, comparable to 80,000–800,000 households, March 19, 2026}.

The workload itself has changed shape. A chat query resolves in 500 milliseconds. An agentic task such as a coding audit, a research synthesis, or a multi-system workflow, resolves in 10 minutes to 14+ hours, and runs in parallel with thousands of siblings inside a single campus. The load profile of an AI data center has converged toward that of an aluminum smelter: continuous, high-wattage, near-100% utilization, 24/7. The bursty inference workload grid planners modeled against is gone. The new profile is baseload.

In addition, agentic compute has a unique memory-power overhead. To maintain state over a 14-hour + task, agents must keep massive context windows active. In 2026, hyperscalers are offloading KV caches to SSD storage to manage the memory bottleneck, but the energy required for constant data shuffling remains a significant hidden cost. Anthropic has published research on “context engineering”. This refers to compaction, structured note-taking, and multi-agent architectures designed to address “context rot” in long-horizon agents^{190Anthropic, Effective context engineering for AI agents: Anthropic describes challenges like “context rot,” “context pollution,” compaction, structured note-taking, and multi-agent architectures for long-horizon tasks, but does not use the term “context entropy,” September 29, 2025}. These techniques reduce the memory-power overhead of maintaining state across extended tasks and indirectly ease the underlying power constraint.

The power infrastructure to support these facilities does not exist in most locations where they are being planned. Grid interconnection timelines run three to five years in the United States, longer in Europe. Permitting for new transmission lines routinely takes a decade.

The US is on track to add a record 86 GW of utility-scale generation in 2026: 51% solar, 28% batteries, 14% wind^{191U.S. Energy Information Administration, New U.S. electric generating capacity expected to reach a record 86 GW in 2026 (51% solar, 28% battery storage, 14% wind; 53 GW added in 2025, the largest annual increase since 2002), April 16, 2026}. Useful firm power near data-center hubs is a fraction of that headline. In PJM, the single most important AI interconnection, utility large-load commitments now exceed accredited generation capacity by 3x^{192American Public Power Association / Wood Mackenzie, Newly added U.S. data center capacity slows down considerably in Q4 2025, with PJM’s utility large-load commitments reaching three times the accredited capacity in its risked generation queue (pre- and post-ISA), March 15, 2026}. China, in contrast, added 434 GW in 2025^{193Yicai Global, "China’s Wind, Solar Power Installed Capacity Exceeded 1,800 GW for First Time in 2025", January 29, 2026} alone and is co-locating compute with generation in the renewables-rich West. This is not a cyclical shortfall the US closes with capex. Rather, it is a profound infrastructure divergence, one that could constitute a Single Point of Failure of the Agentic Era, i.e., the specific vulnerability that, if left unresolved, constrains the entire transition regardless of how fast models improve or how much capital flows into AI. A SPOF, in the framework we introduced in 2025, is the component of a system whose failure compromises the whole structure.

Nuclear is, as of 2026, the only carbon-free source with the density for AI-scale loads, and the scramble for it since 2024 is no longer directional. It is the infrastructure. Microsoft underwrote the restart of Three Mile Island via a 20-year PPA with Constellation (835 MW, 2027-2028)^{194Constellation Energy & U.S. DOE, Constellation is restarting Three Mile Island Unit 1 as the Crane Clean Energy Center under a 20-year PPA with Microsoft, backed by a $1 billion DOE loan toward an estimated $1.6 billion restart cost, with operations possibly resuming as soon as 2027}. Amazon bought Talen’s Susquehanna campus and is developing SMRs with X-Energy and Dominion^{195Data Center Dynamics & PowerMag, AWS buys Talen’s 960 MW Cumulus nuclear campus for $650 million (March 2024) and expands its Susquehanna PPA to 1,920 MW through 2042 (June 2025); Amazon also targets 5 GW of SMRs with X-energy by 2039 and has an MOU with Dominion for a ~300 MW SMR at North Anna}. Google signed Kairos (500 MW SMR fleet) and Fervo geothermal^{196Kairos Power & Fervo, Google–Kairos Master Plant Development Agreement (Oct 14, 2024) targets 500 MW of advanced reactors by 2035 with first unit online by 2030; Fervo’s Project Red (3.5 MW Nevada pilot) scales to a 115 MW NV Energy Clean Transition Tariff project supplying Google’s Nevada data centers, 2021-2026},^{197Power Magazine, Utility Dive, Meta & World Nuclear News, Meta’s January 2026 nuclear deals with Vistra, TerraPower, and Oklo secure up to 6.6 GW of capacity by 2035, January 815, 2026}, Oklo, and Vistra, making it the single largest corporate purchaser of nuclear energy in US history. Sam Altman backed Oklo (2.6% stake, former chair, co-founder of the SPAC that took Oklo public). xAI stood up 200+ MW of on-site gas turbines in Memphis in under a year^{198SemiAnalysis & public records, xAI’s Colossus build-out: Memphis Colossus 1 currently draws roughly 240 MW from gas turbines (a conservative “200+ MW” low-end), with Shelby County’s July 2, 2025 permit covering 15 units totaling 247.2 MW and earlier aerials showing ~35 turbines (~400+ MW nameplate); a joint venture with Solaris Energy in Southaven, Mississippi adds about 400 MW of mobile turbines on the way to ~1.1 GW by 2027, and the initial Colossus complex was erected in roughly 122 days, 2024–2026}. Stargate Abilene is architected as a 1+ GW behind-the-meter campus. Together, these form a coordinated admission that the public grid cannot deliver what agentic compute needs on the timeline required.  Hyperscalers are no longer interconnecting to the grid: they are becoming utilities.

The constraint sharpens as the workload shifts from training to inference. Training runs are episodic. A frontier model trains for weeks in a concentrated burst, then stops. Inference is continuous. Every agent action, every orchestrated workflow, every token consumes power indefinitely, at a load that scales with agent populations.

As the mix shifts, the demand profile moves from periodic spikes to sustained baseload: precisely the pattern that stresses grids most acutely, because it requires always-on generation rather than peak-shaving.

The efficiency gains partially offset this. As noted before, Jevons in this respect operates at industrial scale: cheaper inference makes more workflows viable, expanding total demand even as per-unit consumption falls. Hyperscalers that secure power early through long-term agreements, on-site generation, or strategic location near abundant clean energy acquire advantages that compound. The value of a data center is increasingly determined not by the GPUs it houses but by the megawatts it can draw.

For the AGNT thesis, energy introduces a temporal variable into every projection. The Inference Economy expands at a rate bounded by power availability, regardless of demand or capital. The orchestrators of Parts III to V will find their ability to scale agentic workflows ultimately limited by whether the infrastructure they depend on has the electricity to operate.

The direction is not in question. The speed is. Watts, not weights, now govern the pace of the transition. The pace, as the Two Scenarios in Chapter 18 will argue, is what separates deflationary growth from demand destruction.

§

The Agentic Cloud

The cloud was designed for the SaaS era. The agentic economy demands infrastructure that does not yet exist at scale.

There is a final condition that completes the picture. It is architectural rather than financial or physical. The cloud was built for a different era. SaaS-era cloud was designed for stateless applications, ephemeral compute, hub-and-spoke networking, and consumption-based pricing. An enterprise spins up a virtual machine, runs a workload, and shuts it down. Two decades of perfecting that pattern. It is the wrong pattern for what comes next.

The agentic systems of Chapter 5 generate workloads that differ in every dimension. Every MCP connection, every A2A coordination, every persistent memory retrieval, every multi-step workflow is an inference operation. An agent booking a flight generates 50,000 tokens across research, comparison, booking, and confirmation. This compares to just 500 for a chat query. Agentic workloads multiply token consumption 10x to 100x^{199Iternal Technologies & OpenReview, Token Usage & Cost Projection Guide (2026) and SWE-bench agentic-coding studies: agentic systems typically consume 5–30× more tokens per task than standard chat (ranging from 5,000–15,000 tokens for simple tool-calling up to 1–3.5 million tokens for SWE-bench–style coding), with individual runs sometimes using up to 10× more tokens than others—so illustrative 50,000token, multi-step workflows can reasonably reach 10–100× chat-level token usage, March 2026}.

But the difference is not merely volume. Agents are stateful. They maintain persistent memory and workflow state across sessions, requiring considerably more memory than what traditional workloads demand. The three-tier memory architecture of Chapter 5 requires infrastructure that preserves state, not infrastructure designed to forget:

Agents coordinate laterally. They communicate peer-to-peer through MCP and A2A, requiring sub-100ms latency and dedicated bandwidth between coordinating agents^{200Ehtesham et al., A Survey of Agent Interoperability Protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP), May 2025}. The topology must evolve from a centralized switchboard toward a mesh.

Agents generate explosive storage demands. Vector databases typically experience 10x data expansion, total storage reaching 30x original data after indexing and context persistence^{201DataNorth, The Complete Guide to Vector Databases, April 2026}.

Agents require a different economic model. Pricing must be tied to task completion and outcomes rather than compute hours consumed, which demands infrastructure instrumented for metrics that traditional cloud monitoring was never designed to capture.

These requirements point toward a new infrastructure category: the agentic cloud. Three categories of providers are positioned to build it.

The hyperscalers hold the deepest structural advantages: custom silicon, enterprise relationships built over decades, and captive internal inference demand. Google’s Search, YouTube, Gmail, and Maps are consuming inference continuously, providing a substantial utilization floor regardless of external demand. Their challenge is that hundreds of billions in existing infrastructure were designed for the SaaS era. The organizational complexity of serving legacy and agentic workloads simultaneously is real, and scale alone does not resolve it.

The neoclouds proved something important: that AI workloads demand fundamentally different compute from general-purpose clouds. CoreWeave and other AI-specialist clouds have demonstrated MFU improvements of 15-30% over hyperscaler baselines (e.g., CoreWeave’s 49.2% MFU on H100 training runs vs MosaicML’s 41.85% baseline, per CoreWeave’s published benchmarks). Industry training MFU typically sits in the 35-55% range depending on hardware generation, software stack, and workload. But the neocloud window is narrowing. Their core design is stateless, single-tenant, and hub-and-spoke. This served training-centric workloads well. It is less suited to what agentic systems demand.

The emergent agentic clouds are purpose-built. Daytona Cloud achieves sub-90ms sandbox creation for stateful agent workflows^{202Daytona, Sub90 ms sandbox creation: Daytona Cloud’s agent-native runtime advertises “sandboxes…spinning up in under 90 ms from code to execution,” with internal benchmarks showing ~71 ms sandbox creation and multiple customer case studies (for example, Laude Institute) citing “sub90 ms sandbox creation” as a key benefit, 2024–2026}. Others target observability for multi-agent systems, optimization for coordination-intensive workloads, and outcome-based resource allocation. They carry no architectural debt. They lack scale. They are building for the workload that is arriving.

The model providers see the same opening. Anthropic’s Claude Managed Agents, launched in April 2026, offers the full production stack for agentic workloads as a managed service: sandboxed execution, state persistence, tool orchestration, error recovery, and multi-agent coordination²⁰³. It runs on existing cloud infrastructure, but the developer never touches it. Anthropic abstracts the cloud into a substrate and bills by the session-hour.

The orchestrator’s variable cost is inference tokens. Hardware efficiency sets the floor for what a token costs to produce. The agentic cloud determines what it costs to serve that token inside a coordinated, stateful, production-grade workflow. The companies that optimize across both will set the economic foundation for everything built above them.

§

The Substrate Must Evolve

The AI infrastructure being built in 2026 is the largest private capital deployment in history. It is transforming Big Tech’s balance sheets from capital-light to capital-intensive. It is financed through structures of increasing ambition, from conventional bonds to century paper to securitized GPU assets. And it was committed under assumptions about training workloads that the migration toward inference is already reshaping.

The metamorphosis is substantial but navigable. Inference efficiency is compounding at a pace that makes the unit economics of the agentic economy more favorable with each quarter. The energy constraint governs speed, not direction. And the cloud is beginning, unevenly, incompletely, but perceptibly, to evolve toward the stateful, coordination-grade architecture that agentic workloads demand.

The intelligence is sufficient. The agents are in production. The infrastructure is scaling, and it is starting to change shape. What determines whether specific companies, sectors, and workflows successfully make the crossing is the subject of Part III.

Parts I and II established that the paradigm has shifted. Machines became actors. The intelligence required for autonomous professional work exists as commercially available infrastructure. Agents are deployed in production. The infrastructure, while strained, is scaling.

The question now: Who captures the value it creates?

When a factor of production commoditizes, when intelligence becomes abundant, then value migrates to whoever controls the coordination of the newly abundant input. This migration is a pattern that has repeated across every major economic transition, from electricity to computing to cloud infrastructure. The entity that coordinates the abundant resource captures the surplus. In the agentic economy, the entity that coordinates machine actors captures the surplus of the entire transition.

The framework that describes who captures value, under what conditions, and why is Orchestration Economics. This manifesto defines the Three Laws of Agentic Value that determine whose orchestration position is durable. They are structural in nature, already visible, measurable, and investable.

However, we must first be clear about how the new layer of value forms.

§

Orchestration Economics

This is the framework that describes how value is captured in an economy where intelligence is abundant, labor is partly synthetic, and workflows are no longer coordinated by humans moving between tools but by agents moving between systems on behalf of humans.

Orchestration Economics rests on a fundamental structural claim: The entity that controls the orchestration of machine actors captures the surplus of the transition. This is not a marginal improvement over SaaS economics, nor is it limited to the field of software. It is a categorical shift in three dimensions:

Pricing is outcome-based, not seat-based, or time-based. The unit of commerce is a claim processed, an analysis completed, a route optimized – not a license sold or a billable hour. Pricing per agent task (i.e., per consumption) is an intermediary step in that journey.

Budget is labor, not IT. The vendor does not compete with software for a slice of the 3-5% IT line. It competes with salaried professionals for a slice of the 25-40%^{204Ramp, "Selling, General & Administrative Expenses (SG&A): Definition & Examples"} knowledge-work line. The addressable market expands by a factor of 5-10x. What is more, the market is not limited to labor substitution with agents completing value accretive tasks that did not exist before.

Defensibility is topological, not feature-based. What protects an orchestrator is the structure of the workflow it controls, not the capabilities of any single agent inside it.

The rest of this chapter establishes where this layer sits in the new architecture of the firm, how value migrates into it, why most companies will fail to capture it even when they think they are, and under what conditions an orchestration position is durable rather than temporary. The Three Laws formalize those conditions.

§

The New Layer

Before agents, humans went to tools. Now, tools come to humans, mediated by an agent that sits between intent and execution.

Before agents, the architecture of knowledge work was straightforward. A human expressed intent: process this claim, close this deal, optimize this route. The human also had the tools to execute that intent. The human opened a CRM to update a record, a spreadsheet to run an analysis, and an email client to send a message. Each tool was a destination. The human was the orchestrator. The tools were the instruments. The business outcome was the product of human judgment applied through software.

Agents invert this architecture fundamentally. They insert a new layer between the human and the tools. The human used to go to the tools. Now the tools come to the human, mediated by the agent.

In the simplest case, this involves just a single AI agent. The human expresses intent: update the Chicago deal to closed-won and schedule a kickoff with their team next week”. The agent receives that intent, determines which systems to invoke (the CRM, the calendar, the email system), executes the necessary actions across those systems, and delivers the outcome. The human never opens the CRM. Never touches the calendar. Never drafts the email. The agent coordinates it.

Even in this single-agent case, orchestration is happening. The agent is orchestrating tools on the human’s behalf. The agent is, in the precise sense of the word, a synthetic colleague. This is the foundation of Orchestration Economics: agents are not better tools. They are a new layer between humans and their tools, designed to produce business outcomes.

The entity that controls the “Orchestration Layer” owns the relationship between the human and the work for a specific workflow, or set of workflows, aimed at producing a business outcome. It decides which tools get invoked, which data gets accessed, which sequence produces the best outcome, which systems are essential, and which are bypassed entirely.

The Orchestration Layer that this Manifesto describes is categorically different from the technical Orchestration Layer known as the “Harness”. This is a critical distinction. The Harness allows an LLM to produce a result and makes intelligence operational. The Orchestration Layer turns operational intelligence into a business outcome.

§

The Architecture of the Agentic Firm

The Agentic Era company is not monolithic. It has three concentric sub-layers or “rings”.

The agentic firm is organized in concentric rings, not layers, because what matters is not just function, but proximity to value capture. The outer the ring, the greater the control over outcomes, and the stronger the economic position:

Ring 1 – Intelligence. The foundation models. The substrate. Its economic unit is the token.

Ring 2 – Harness. The technical orchestration infrastructure that makes intelligence operational: the systems that decompose goals, delegate to agents, manage state, coordinate execution, and accumulate learning across sessions. Claude Code, Cowork, Managed Agents, OpenAI’s Frontier, and Codex are Harness products. Its economic unit is the session.

Ring 3 – Orchestration. The outermost sub-layer. The position occupied by the entity that places Rings 1 and 2 at the center of its operations and adds what neither can produce: the operational context accumulated through running the world. Its economic unit is the outcome.

When I first argued in 2025 that the durable moat for LLM labs lay in orchestration rather than model intelligence, I was describing Ring 2. The labs confirmed the thesis. They recognized that model capability alone was commoditizing and moved into the infrastructure that makes it operational. The Coding Wedge was the entry point. Each product since has deepened the position.

But the labs did not stop at the Harness. Cowork extended into knowledge work. Vertical plugins reached into business functions. Frontier positioned itself as a control plane for the enterprise. With each step the labs move outward, from intelligence through Harness toward the outermost sub-layer, contesting it against companies that have occupied it for decades. This expansion from Ring 2 to Ring 3 is at the root of the SaaSpocalypse and the broader value destruction in the public markets since the start of the year. We explore this in Chapter 13 infra.

An entity that provides intelligence is not an orchestrator. An entity that builds a Harness is not an orchestrator. The orchestrator wraps around both and produces business outcomes.

A single entity can occupy more than one ring. Two patterns are already visible.

The labs want to occupy all three. Anthropic today credibly sits in Ring 1 (Claude), Ring 2 (Claude Code, MCP, Skills, Agent SDK), and extends into Ring 3 (Claude Code for enterprise, Cowork, managed agents). In narrow domains, coding is the clearest because it already crosses all three. OpenAI is pursuing the same stack. The labs built inward-out: from intelligence, through Harness, toward outcome.

Occupying the inner rings does not confer the outer ring. Ring 3 is not a further technical layer. It is operational context co-produced with the workflow being orchestrated. The labs can ship Ring 2 unilaterally. They cannot ship Ring 3 unilaterally, because the intent, the context, and the workflow live inside the firm whose work is being orchestrated.

In domains where the lab also owns the workflow, such as coding, where the developer’s intent, context, and workflow are all captured natively inside the Harness, the lab credibly holds Ring 3. In domains where it does not, such as insurance underwriting at a carrier, credit analysis at a bank, claim evaluation, or at a reinsurer, then the lab supplies Rings 1 and 2, and someone else holds Ring 3.

Ring 3 is, therefore, domain-specific, not firm-wide. The same entity can hold it in one workflow and lose it in another. Whether any entity durably holds Ring 3 is not settled by its stack position. It is determined by the Three Laws.

The architecture is new. The contest for who occupies the outermost ring is already underway.

§

Captain and Crew: The Topology of Coordination

Orchestration is also a technical-architectural fact about how multi-agent systems produce coherent outcomes. The topology that wins is centralized, not federated. That topology is where value accrues.

In a single-agent system, the agent orchestrates tools. In a multi-agent system, the orchestrator orchestrates agents who orchestrate tools. The structural position is the same. The layer just becomes thicker, more complex, and more valuable.

For a simple task such as updating a CRM record or scheduling a meeting, a single agent coordinating a few tools is sufficient. But to evaluate an insurance claim, analyze a company’s creditworthiness, or optimize this supply chain, the task becomes too complex for a single agent to handle every dimension.

This is where multi-agent systems emerge. The system has a crew of agents with specialized capabilities whose job is to execute well-scoped work: document extraction, risk modelling, fraud detection, pricing, or compliance validation. They can be brilliant within their domains. But they do not see the whole map. They do not decide which workflow to run. They do not own the relationship with the human who asked for the outcome.

A multi-agent system could, in principle, be organized in many ways. A federation of peer agents could negotiate amongst themselves. A swarm could converge by consensus. Independent specialists could run in parallel and average their outputs. All of these have been tried. None produce coherent work at enterprise reliability.

The 2025 Google-DeepMind-MIT study formalized this structure. A multi-agent system is an agent system with more than one reasoning entity, where agents interact through a communication topology and an orchestration policy that defines how the system makes decisions: how sub-agent outputs are aggregated, whether the orchestrator can override sub-agent decisions, whether memory persists across rounds, and when the task is complete.

As Chapter 5 established empirically, centralized coordination reduces error propagation by roughly 75%. This result emerges from the physics of coordination, not from any individual agent’s capability. Independent topologies amplify errors by 17×. Decentralized debate catches some but imposes an 8× overhead of its own. Hybrid topologies come in at 5×. Centralized at 4×74. Only the architecture where one agent receives the task, decomposes it, selects and validates specialists, and decides what happens next produces the coherent outcomes enterprises will pay for.

The orchestrating agent becomes the captain. It does not perform the specialized work itself. It decomposes the goal into subtasks. It assigns those subtasks to the appropriate specialists on the agentic crew. It collects their outputs. It validates them. It synthesizes the results. The human talks to the captain. The captain commands the crew. The crew does the work. The captain reports back to the human. That is the agentic AI loop.

As we move into the core exploration of Orchestration Economics, it is worth noting that the discussion about the passage into the Agentic Era has remained at a simplistic, binary stage. Winners and losers. Eat or be eaten. Victim or beneficiary of the SaaSpocalypse. The more nuanced and challenging question most companies will be facing: Do you want to be the captain or the crew?

The economic consequence follows directly from the topology, not from the metaphor.

Specialist agents are substitutable. A better extraction model can be swapped in. A better pricing model can replace the incumbent. The crew is upgradeable. Its members have limited pricing power because the captain can always source a better one.

The captain is not substitutable. Replacing the orchestrator means rewiring the workflow, including remapping the topology, retraining the policy, reestablishing every integration, and re-accumulating the operational memory. The switching cost is categorical.

Value is multiplicative, not additive. A single agent coordinating three tools produces linear value: the sum of the tasks automated. An orchestrator coordinating five specialist agents, each coordinating their own tools, produces nonlinear value. The orchestrator can redesign the workflow, parallelizing steps, eliminating redundancies, and catching errors across domains that no individual specialist would detect. The orchestrator is not a participant in the workflow. It is the workflow.

In human organizations, we call this management. In multi-agent systems, we call it orchestration. Value accrues to the Orchestration Layer rather than distributing evenly across the agents beneath it. Not because someone has to be in charge, but because the topology in which centralized coordination wins is the only topology in which the workflow works at all. The entity at the center of that topology captures the surplus that coherence creates.

Let’s follow the money through an actual enterprise workflow:

A commercial insurance submission arrives. Under the old architecture, it enters a human chain: intake, review, coding, investigation, pricing, approval, and documentation. Each human is an expert. Each has trained for years. Together, they process the claim in days, sometimes weeks, at costs measured in hundreds of dollars per claim. The Orchestration Layer lived in a manager’s brain and a project management tool. The bottleneck was human capacity to coordinate.

Under agentic orchestration, the same workflow transforms. The submission arrives at a master agent. The master agent does not “process a document” in the manner of legacy software. The master agent interprets a goal: evaluate this risk and produce a quote that fits our growth and risk constraints. It inspects the case. It decides which specialized agents to invoke: extraction, enrichment, risk modeling, pricing, or compliance validation. It sets their inputs. It sequences their work. It validates their outputs. It determines when the goal is achieved.

The specialist agents are valuable. They have real reasoning capacity. But they are no longer where the workflow is decided. Their job is to respond to the plan. Not to define it. The control plane has moved. So has the economics.

The question every company in the agentic economy must answer is not rhetorical. It is operational.

Are you the captain? Or are you the crew?

§

The Programmable Firm

The addressable market of Orchestration Economics is not set by what AI can do. It is set by what work can be programmed.

Carnegie Mellon and Stanford University research teams released a landmark study in October 2025 called “How Do AI Agents Do Human Work?”  that provided an empirical foundation for the Agentic Era. The teams compared how humans and AI agents complete identical work across data analysis, engineering, computation, writing, and design. They concluded that for tasks defined as “readily programmable”, AI agents will complete the work 88.3% faster at 90.4-96.2% lower cost^{205Zora Zhiruo Wang, Yijia Shao, Omar Shaikh, Daniel Fried, Graham Neubig, Diyi Yang, "How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations", Carnegie Mellon University & Stanford University, arXiv:2510.22780, October 26, 2025}.

The key finding is not that work is programmable end-to-end. It is more consequential: programming is embedded across 82.5% of knowledge work occupations. In practice, this means a large share of enterprise workflows contain programmable components, even if they are not fully automatable.

Programmability is the unlock. A workflow that can be programmed can be agentized. A workflow that can be agentized can be orchestrated. A workflow that can be orchestrated can be sold as an outcome rather than licensed as a tool. The economic category changes not because the AI is better, but because the work itself was always more structured than the software market reflected.

The gate is reliability. Every enterprise workflow has a reliability requirement: a level of accuracy and completeness below which human oversight remains necessary. When an AI system handles 40% or 50% of cases correctly, it accelerates human workers but cannot replace them. Every output still needs review. The human chain stays intact. The company has purchased a productivity tool.

When an AI system handles 85% or 90% of cases correctly, the economics change categorically. The workflow runs autonomously. Humans handle only the exceptions, the fraction of cases the system flags for review. The company has not purchased a productivity tool. It has purchased an autonomous function. The difference between these two states is not a percentage improvement. It is a reclassification.

In the first state where AI is a productivity tool, the vendor sells software licenses. The buyer pays $50-200 per seat per month from the IT budget. The vendor competes on features against other software vendors. The human labor cost remains on the books. This is SaaS economics.

In the second state where AI becomes an autonomous function, the vendor sells an outcome. The buyer pays per claim processed, per credit analysis completed, or per supply chain route optimized. The budget is no longer IT spend. It is labor spend: the fully loaded cost of the knowledge workers the autonomous function replaces.

A single insurance underwriter costs $8,000-15,000 per month in salary, benefits, training, and management overhead. A credit analyst costs similar. The vendor prices against that labor cost, not against competing software licenses.

Apply this lens to a large institution like JPMorgan. With approximately $50 billion^{206JPMorgan Chase & Co., 2025 Annual Report (Form 10-K): total compensation expense of ~$50–51 billion for full-year 2025; 318,512 employees as of December 31, 2025. Filed with the SEC, February 2026} in annual labor costs and assuming ~30% of tasks fall into the “readily programmable” category, agents begin to meaningfully reshape the cost structure. However, because these tasks still require verification, exception handling, and integration into broader workflows, the immediate impact is not full replacement but partial automation and labor reallocation. Even under conservative assumptions, this translates into multi-billion-dollar efficiency gains, not through elimination of labor, but through compression of the programmable portions of work. That pool does not sit in the IT line. It sits in headcount.

The addressable market shifts accordingly. Enterprises typically spend 3-5% of revenue on IT, but 25-40% on knowledge work. The critical transition is not from software to AI, but from tools to workflow execution. Once an orchestrator reliably automates the programmable segments of work while coordinating human oversight where needed, it begins to tap into the much larger pool of labor spend. The prize is therefore not replacing SaaS but expanding beyond it by capturing portions of the operating budget historically reserved for human execution. In that sense, the orchestrator that crosses the reliability threshold does not compete within the software market but instead redefines the boundary of the market, expanding it by a multiple of five to ten.

The orchestrator creates this transition. The specialist agents, no matter how capable individually, do not. A pricing agent that achieves extraordinary accuracy still operates within the first state. It makes the human underwriter faster. It is a tool.

An orchestrator that coordinates pricing, risk, compliance, and documentation agents into a coherent autonomous workflow delivers the second state. It replaces the underwriting function. It delivers an outcome. The outcome emerges only from the coordinated whole. Only the orchestrator delivers it.

§

Bypass Risk

If an agent intercepts intent before it reaches your system and routes it to execution without invoking you, your position evaporates. You become the crew on someone else’s ship.

Orchestration Economics does not guarantee permanent advantage. The same structural logic that creates the orchestrator’s position also defines the conditions under which that position can be undermined.

Bypass risk is the possibility that a new entrant positions itself between the user’s intent and the existing orchestrator. In doing so, it captures the Orchestration Layer from above. Protocols such as MCP standardized how agents connect to external systems and to each other. A2A standardizes agent-to-agent communication. ANP sits on top of both as a discovery layer.

Standardization is powerful for interoperability. It also dissolves the integration friction that protected many orchestration positions. The company that controlled its workflow through the sheer difficulty of connecting systems wakes up one morning to find that friction gone. Protocols dissolved it. Orchestration Economics is therefore not a simple claim that the orchestrator always wins. It is a framework that says the orchestrator wins unless it gets bypassed.

The orchestrator’s advantage is structural, but its defensibility is conditional. It depends on building moats that protocols cannot standardize away and that competitors cannot replicate. Which moats? Under what conditions does an orchestration position become durable rather than temporary? What separates the orchestrators who compound their advantage from those who lose it to the next layer above?

§

The Three Laws

Knowing that orchestrators capture value does not tell us which positions are defensible. The orchestrator’s advantage is structural. Its defensibility is conditional. It depends on building moats that protocols cannot standardize away and that competitors cannot replicate.

Three conditions determine whether an orchestration position is durable. Each is necessary. None is sufficient alone. The entity that satisfies all three simultaneously holds a position that compounds with use, resists bypass and captures the surplus of the agentic transition. The entity that satisfies one or two, even brilliantly, occupies a layer that the orchestration economy will eventually treat as crew rather than captain.

Chapter 8 develops the First Law: Proximity to Intent Determines Value Capture.

Chapter 9 develops the Second Law: Context Builds Moats.

Chapter 10 develops the Third Law: Workflow Intelligence Secures Control.

Once we define the Three Laws, we will apply them to Palantir.

The entity closest to the moment a human expresses a goal captures routing authority, relationship ownership, and the economics of the phase transition. Everything downstream is an API call.

§

The Intent Hierarchy

Agents operate across the Orchestration Layer, but their economic position depends on their proximity to the moment a human expresses intent.

Let’s return to the insurance workflow. The master agent receives a goal: evaluate this risk and produce a quote that fits our growth and risk constraints. The extraction agent receives a task: pull the relevant data from this document. The fraud detection agent receives a task: check this submission for anomalies. The pricing agent receives a task: calculate the payout given these parameters.

All four are agents. All four orchestrate tools. All four sit in the Orchestration Layer. But the master agent controls routing. It decides which specialists to invoke, in what order, and under what constraints. It defines the workflow. It validates the outputs. It determines when the goal is achieved. The specialist agents execute tasks that someone else defined. They may be brilliant within their domains. But they do not see the whole map. They do not decide whether the claim gets processed. They do not own the relationship with the human who submitted it.

Chapter 7 established that this hierarchy creates disproportionate value capture. The First Law explains why: it is not coordination alone that creates the advantage. It is proximity to the human’s original intent. The master agent sits closest to the moment a human expressed a goal. The pricing agent sits furthest from it, several translations downstream. That distance determines everything: defensibility, pricing power, switching costs, and bypass risk.

The principle generalizes across the entire agentic economy. There is a hierarchy of positions relative to intent, and your location determines the durability of your orchestration position:

Intent Origin carries the highest value. This is where goals are first expressed in natural language, unstructured, with high ambiguity. The human who submits “evaluate this risk” is expressing intent at the origin. In enterprise software more broadly, search interfaces occupy this layer: Google when you type a query, ChatGPT when you express a goal, Slack when you message a team channel about a problem. Whoever controls intent origin potentially controls routing.

Intent Capture carries strong value. This is where goals become structured as they are translated into specific actions, mapped to known workflows, and scoped with constraints. When the master agent decomposes “evaluate this risk” into discrete tasks for each specialist, it is performing intent capture. If agents intercept intent at the origin layer and route it directly to execution, the capture layer is skipped entirely.

Intent Processing carries moderate value. This is where goals execute. Data is read and written, APIs are called, and workflows run. The extraction agent and the fraud detection agent occupy this layer. Essential work. Nothing happens without it. But it does not control where intent originates or how it gets structured.

Intent Output carries the lowest value. This is where results are viewed as dashboards refresh, reports generate, and notifications display. This layer is one abstraction from replacement. When the master agent synthesizes results and delivers the outcome directly to the human who expressed the goal, separate output displays become redundant.

The further upstream you sit, the more durable your moat. The further downstream, the more replaceable you become. Because the Orchestration Layer above you can substitute you without the user noticing. The insurance company does not care which extraction agent runs. It cares that the master agent delivered the right quote.

§

Why Upstream Wins

The entity at intent origin has routing authority, relationship ownership, and phase transition capture. Entities downstream have none of these. The asymmetry is structural.

The hierarchy is not merely a ranking. It reflects a structural asymmetry in how value flows through agentic systems. The entity at intent origin has three advantages that entities downstream do not:

Routing authority. The intent origin decides which downstream systems get invoked. Every goal expressed is an opportunity to engage your capabilities, route to your ecosystem, monetize through your platform. This is the economic engine that made AWS successful: control the compute layer, and you can upsell storage, databases, machine learning services, and a thousand other capabilities. Intent capture is the agentic equivalent. Control the orchestration surface, and every user goal becomes a distribution channel.

Relationship ownership. The human interacts with the intent surface. The human trusts the intent surface. The human returns to the intent surface. Everything downstream is invisible to the user. It is mediated, abstracted, and replaceable. The intent surface owns the relationship. Everyone else is a vendor.

Phase transition capture. Connect this to the economics of Chapter 7. The phase transition to Orchestration Economics only accrues to the entity that delivers the complete outcome. That entity must sit at or near intent origin, because only intent origin can promise the customer that if receives the goal it will deliver the result. An orchestrator at the processing layer can coordinate agents effectively, but it cannot deliver the autonomous function the customer is paying for. It depends on someone upstream to provide the goal. Someone upstream can bypass it. An orchestrator at intent origin owns the goal itself, and everything downstream flows from that ownership.

Not all data enables these advantages equally. Customer behavioral data, such as how users interact, what they choose, what patterns they follow, creates the strongest orchestration advantage because it feeds directly into routing decisions. Operational transaction data creates the strongest execution advantage but does not grant routing authority. The implication for investors is direct: companies that accumulate behavioral data at the intent surface compound their routing advantage with every interaction. Companies that accumulate transactional data downstream build strong specialist positions but remain vulnerable to bypass. The corollary is severe for companies positioned downstream. If you are a point solution, such as a specialized sales engagement tool, and the user’s goal is expressed in Slack and orchestrated by Agentforce, you are competing to be the sub-agent that gets invoked. Your switching costs collapsed. Your direct relationship with the user eroded. You moved from captain to crew. The crew is replaceable.

§

The Software Inversion

When an agent mediates the interaction, the system of record becomes just another API call. Moats no longer form where data sits. They form where intent originates.

For two decades, enterprise software moats were constructed downstream. You won by owning the system of record. This is the database where critical business data lived, surrounded by workflows that made switching painful. Salesforce built an empire on this principle. So did Workday, ServiceNow, and SAP.

The Orchestration Layer inverts this. When an agent mediates the interaction, the system of record becomes just another API call. The agent does not care whether your CRM is Salesforce or HubSpot. What matters is whether the agent can access the right context and orchestrate the sequence of actions needed to fulfill the user’s goal. Moats no longer form where data sits. They form where intent originates.

Salesforce recognized this before most of the market. Its repositioning of Slack at Dreamforce 2025^{207CX Today, Dreamforce 2025: top announcements ft. Agentforce 360, the new Slack, Apromore acquisition, 2025} is the First Law made concrete. Restricting AI companies’ access to Slack data looked defensive. The sequence that followed^{208Salesforce Ben, 3 Slack announcements from Dreamforce 25: What’s new, 2025} instead demonstrated an offensive strategy by governing access through MCP, the Agentforce runtime, and first-party agents embedded directly in Slack.

Salesforce moved the moat upstream, from owning the database where records sit to owning the surface where goals are expressed. Slack became the system of intent. Customer 360 remained the system of record. Agentforce became the Orchestration Layer binding them. The company understood that controlling intent origin is more valuable than controlling any system downstream.

The same principle operates at every scale. Within OpenAI’s GPT-5, a built-in router analyzes user intent and directs processing to the appropriate internal model variant. The router sits at intent origin inside the model. Platforms like Slack or Teams must do the same thing across entire ecosystems. The architecture is fractal: whoever captures intent first controls what happens next, whether “next” means selecting a model variant or orchestrating a fleet of enterprise agents.

§

The Protocol Test

Protocols can standardize every layer below intent origin. They cannot standardize the moment a human expresses a goal. Intent origin is protocol-resistant because it depends on human habit, institutional workflow, and accumulated context, none of which can be specified in an API schema.

Protocols provide the sharpest test of the First Law.

Chapter 7 established bypass risk as the counterforce to orchestration advantage. MCP, A2A^{209Linux Foundation, "Linux Foundation Launches the Agent2Agent Protocol Project," June 23, 2025}, and the rapid adoption of both by every major platform standardize how agents connect to systems and to each other. That standardization dissolves integration friction, the difficulty of connecting systems that once protected many orchestration positions. The First Law predicts exactly which positions survive this dissolution and which do not.

Protocols can standardize every layer below intent origin. They can make tools interchangeable, specialist agents substitutable, and integrations frictionless. If every CRM exposes the same agent endpoints via MCP, an orchestrator has little structural reason to prefer one over another. Switching costs collapse. Differentiation shifts toward price. Margin compresses. This is the crew’s fate in a world of universal protocols.But protocols cannot standardize the moment a human expresses a goal. They cannot standardize the trust relationship between a user and the surface where they articulate what they want. Intent origin is protocol-resistant because it depends on human habit, institutional workflow, and accumulated context. None of these can be specified in an API schema.

Protocols commoditize execution. Intent surfaces capture value.

The orchestrator that relied on integration friction has no moat left. The orchestrator at intent origin has a moat that protocols reinforce because every new protocol-compliant tool that connects to its ecosystem makes the intent surface more valuable, not less. More tools mean more capabilities the captain can route to. The crew gets more substitutable. The captain gets more powerful.

§

Strategic Positioning

If you neither own intent nor provide differentiated capabilities to the orchestrator, you risk becoming a replaceable API call in someone else’s plan.

The First Law creates a simple diagnostic for any company in our investment universe: where do you sit in the intent hierarchy, and can you move upstream?

For the systems of record, the processing engines, the output layers, the imperative is to move from downstream to upstream. Build the agent cockpit, not just the system of record. If moving upstream is not feasible, differentiate so aggressively that you become irreplaceable to orchestrators. Be the specialist so good that the captain cannot sail without you. Workflow lock-in will not protect you. Data gravity will not protect you. The question is whether you can become the agent that other agents depend on.

For horizontal orchestrators competing to be the default intent surface, the imperative is speed. Race to become the place where users start their day and express goals. This probably means conversational interfaces rather than traditional application UIs. Invest in context accumulation and memory architectures, because the depth of your conversational memory is what prevents bypass from the next intent surface up.

For specialized agents at the execution layer, embrace protocols and compete on quality, speed, and cost. Be the best sub-agent in your category, and you will be routed to millions of times. Fail to be best-in-class, and you will be replaced overnight. You are crew. Make yourself indispensable crew.

Where does value concentrate five years from now? Three scenarios frame the range:

Fragmented orchestration: every application embeds its own agent layer. Slack has Agentforce. Microsoft has Copilot. Intent capture is distributed, and moats remain traditional.

Consolidated orchestrators: a handful of platforms emerge as dominant intent gateways, orchestrating best-of-breed sub-agents behind the scenes. Downstream tools become commoditized utilities.

Decentralized agent mesh: open protocols enable any agent to invoke any other agent, and value flows dynamically based on performance, not ownership.

History suggests a hybrid: consolidation around dominant orchestrators, plus a long tail of specialized agents accessible via protocols. But in all three scenarios, proximity to user intent is the strategic high ground. The players who control where goals are expressed will have structural advantages in routing, context access, and monetization. The players who do not will compete on price.

If you neither own intent nor provide differentiated capabilities to the orchestrator, you risk becoming a replaceable API call in someone else’s plan.

§

The First Law

Own the intent. Orchestrate the work. Everything else is an API call.

The First Law establishes the first condition for defensible orchestration: proximity to intent determines where value concentrates, who controls routing, and whose position survives protocol standardization. Moats form where intent originates, not where data sits. The captain who hears the human’s goal first captures the economics of the phase transition. Everyone downstream competes on execution.

Owning intent is necessary, not sufficient. Return to the insurance example one more time. The master agent sits at intent origin. It controls routing. It captures the phase transition economics. But what happens when a competitor builds a rival master agent for the same workflow? Both sit at intent origin. Both offer orchestration. What determines which one the customer stays with?

The answer is not who arrived first. It is who built the deeper context. Who accumulated the operational intelligence that makes orchestration effective in ways a new entrant cannot replicate on day one?

That is the Second Law.

Only an agent with deep operational context produces reliable outcomes.

§

Two Master Agents

The First Law is necessary but not sufficient. Two orchestrators can sit at the same point in the intent hierarchy and produce wildly different outcomes. What separates them is invisible on an architecture diagram.

Two insurance companies deploy master agents. Both sit at intent origin. Both orchestrate specialist agents for extraction, risk modeling, fraud detection, and pricing. Both have crossed the reliability threshold and operate as autonomous functions. Both charge per outcome rather than per seat.

On day one, they are equivalent. By month six, one is pulling away. By year two, the gap is too wide to close. Not because it has better models. Both license the same foundation models. Not because it has more data. Both process similar volumes. Because it has deeper context. And context, unlike data, compounds.

This is the Second Law: In the agentic economy, competitive advantage belongs to whoever accumulates the richest operational context and embeds it in their Orchestration Layer. Operational context becomes the “soil” for agents that ultimately allow to achieve business outcomes.

Context is not data. It encompasses data when built correctly through sophisticated memory architectures that layer behavioral patterns, relationships, and temporal understanding into something that makes agents genuinely intelligent.

Data tells you what happened. Context tells the agent what to do next.

§

Why Context, Not Data

Enterprise AI is discovering a painful truth: data alone does not create intelligence. What separates systems that stall in pilot purgatory from those that deliver real outcomes is not the volume of information they possess, but the context accumulated through sustained operational experience.

The enterprise software industry spent the last decade obsessed with accumulating data. Companies built vast data lakes and warehouses, convinced that whoever collected the most information would dominate their markets. That conviction is now failing its most important test.

The MIT GenAI^{210Aditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari (MIT NANDA), "The GenAI Divide: State of AI in Business 2025," July 2025} study released in July 2025 caused such industry uproar by reporting that 95% of corporate GenAI projects failed to get past the pilot stage. This was widely misinterpreted as a judgment of the inherent limitations of generative AI. What most readers missed was one of the key reasons researchers identified for these failures: model quality fails without context. The models had access to petabytes of enterprise data. What they lacked was the operational intelligence that transforms data into effective action.

The distinction is fundamental. Data is static. It sits in databases. It can be queried, copied, or transferred. It answers the question “what happened?”.

Context is dynamic. It incorporates behavioral patterns, relational understanding, temporal awareness, and causal reasoning that emerge from sustained operational interaction. It answers the question “what should happen next, given everything we know about this specific situation?”.

Return to the two insurance companies. Both have the same claims data in the form of submission histories, payout records, and policy details. But only one has accumulated context: the knowledge that claims from certain regions involving certain diagnoses follow a predictable escalation pattern, that this particular adjuster’s flag for fraud has a 94% confirmation rate while that one’s has a 61% rate (illustrative numbers), that Q4 submissions from mid-size manufacturers spike in complexity because of year-end inventory adjustments.

This is not data in a table. It is operational intelligence that emerged from thousands of orchestrated interactions, encoded in the memory architecture of the system.

The master agent with this context makes better routing decisions. It assigns the right specialist to the right claim. It anticipates complications before they manifest. It produces more accurate outcomes faster. The master agent without it has the same capabilities but operates blindly. It is technically competent but operationally naive.

§

The Four Dimensions of Context

One company knows what you did last month. Another knows what you will do next Thursday evening, when you are tired but not rushed, based on forty-seven matched occasions along sixty-three calibrated dimensions. A new entrant can build equivalent features. It cannot manufacture forty-seven matched occasions.

Context is not a single thing. It is multidimensional, and the dimensions interact in ways that create compound advantage:

Behavioral context captures how users and processes actually operate as opposed to how they are documented. An orchestration platform tracking how underwriters navigate workflows, such as which shortcuts they develop, where they encounter friction, and how their patterns evolve, accumulates understanding that informs agent optimization. Users claim to follow the manual. The behavioral context reveals what they actually do. One company knows what you did last month. Another knows what you will do next Thursday evening, when you are tired but not rushed, based on forty-seven matched occasions along sixty-three calibrated dimensions. That depth cannot be replicated from scratch. Those occasions accumulate only through observation over time.

Transactional context encompasses decision history and outcome tracking across extended time frames. The master agent that has seen how a thousand similar claims were resolved, including which were escalated, which approved automatically, and which triggered fraud investigations that proved correct, operates in a different cognitive universe than one processing its first hundred. A procurement agent accessing complete transaction histories makes recommendations informed by actual results rather than static supplier profiles. The accumulation over years creates understanding that new entrants cannot replicate without a similar operational history.

Relational context maps connections and causal relationships that explain why patterns emerge. Understanding that suppliers deliver reliably for certain categories but struggle with others, that certain customer segments respond differently to pricing strategies, that workflow bottlenecks stem from interdependencies between teams rather than individual inefficiency. This allows agents to reason about causality rather than correlation. The distinction matters for autonomous decision-making, where agents must predict how actions cascade through complex systems.

Temporal context captures time-series patterns and seasonality. A manufacturing orchestration agent that understands seasonal demand fluctuations, maintenance cycles, and capacity constraints makes dramatically different decisions than one operating on point-in-time data. The agent that acts before conditions change outperforms the agent that reacts after.

There is a fifth dimension the framework must account for, one that emerged from the first production systems operating at enterprise scale. OpenAI’s internal data agent^{211OpenAI, Inside OpenAI’s in-house data agent, January 29, 2026; VentureBeat, OpenAI’s AI data agent, built by two engineers, now serves thousands of employees, March 2026; The New Stack, Kepler: OpenAI’s Internal Agent Platform for Synthesizing Data, January 5, 2026} – built to reason across 600 petabytes and 70,000 datasets – discovered that schema metadata and operational history were insufficient. Two tables can look identical at the structural level but differ in ways only the pipeline code that produced them reveals: one includes logged-out users, the other does not; one captures first-party traffic, the other captures everything. OpenAI had to crawl its own codebase to teach the agent what the data means, not just what it contains. Constructive context – the understanding of how institutional data was built, filtered, transformed, and aggregated – prevents the class of errors that arises when an agent interprets a field correctly at the schema level but wrongly at the business logic level. Without it, the agent produces technically correct but institutionally wrong answers.

This is the reason Informatica has spent seven years building CLAIRE, a metadata knowledge graph that maps lineage, governance policies, and semantic relationships across fragmented enterprise data – compressing what was once a week-long schema mapping exercise into minutes and exposing, via MCP, the constructive context that agents need to reason correctly over enterprise systems^{212Informatica CLAIRE and metadata knowledge graph; auto-mapping GA in Summer 2025 release. MCP support enables external AI agents to access CLAIRE’s metadata intelligence while maintaining governance and compliance. Informatica acquired by Salesforce, 2025}. It is also the mechanism visible in the Claude Code source architecture described in Chapter 18: the memory consolidation system that rewrites its own index – resolving contradictions, pruning stale facts, converting relative references to absolute – is not storing data. It is building constructive context from operational reality, session by session, so that the next interaction begins from institutional understanding rather than raw observation^{213Claude Code v2.1.88 "Dream" consolidation system. See Chapter 18, "The Orchestration Graph.”}.

Each dimension is valuable on its own. Together, they produce something qualitatively different: an agent that understands its operational reality the way an experienced human operator does. Except it never forgets, never transfers to a competitor, and improves with every interaction.

§

The Flywheel

The logistics carrier that has routed a billion shipments, the insurer that has processed ten million claims, and the bank that has underwritten a trillion dollars in credit. Each possesses contextual foundations that no amount of compute can synthesize. The context is a byproduct of operation. You cannot buy it. You earn it.

Better context produces better agent performance. Better performance attracts more users and deeper platform integration. Deeper integration generates richer context. The cycle accelerates. The companies establishing early leads in context accumulation open gaps that widen with every transaction.

This is why incumbent enterprises in regulated or physically grounded industries possess advantages that their technology adoption speeds obscure.

Their operational context, accumulated over decades, provides foundations that digital-native competitors cannot bootstrap overnight. The logistics carrier that has routed a billion shipments, the insurer that has processed ten million claims, the bank that has underwritten a trillion dollars in credit. They each possess contextual foundations that no amount of compute can synthesize. The context is a byproduct of operation. You cannot buy it. You earn it through sustained execution.

§

From Data Moats to Memory Architecture

Recording is storage. Memory is something else entirely. Memory connects past to present in ways that shape the future. The systems competing for orchestration position are not building better databases. They are building minds.

The previous era’s moat was the database. This system of record surrounded by workflows made switching painful. Context moats require a different architecture: memory systems that learn, not databases that store.

The distinction is not semantic. Recording is storage. Data enters, persists, and can be retrieved. Every database does this. Recording is table stakes. Memory is something else. Memory connects past to present in ways that shape the future. It learns from what happened. It updates its understanding. It changes behavior based on accumulated experience. In the human mind, memory is not a filing cabinet. It is a living process, constantly reconstructing the past to predict tomorrow. The systems competing for orchestration position are building memory architectures that function similarly.

The engineering required is substantial. Implementing hierarchical memory with distinct modules for storage, updating, retrieval, and generation requires infrastructure that extends well beyond traditional database capabilities. Short-term memory maintains conversation and task context. Mid-term memory captures session-level patterns and user preferences. Long-term memory preserves persistent knowledge across extended timeframes. That includes user profiles, organizational patterns, and domain expertise. Each tier requires different update mechanisms, different retrieval strategies, and different governance policies.

As UCL and Huawei’s Noah’s Ark Lab demonstrated through Memento126, agents can achieve state-of-the-art performance through external memory without retraining the underlying model. They do this by storing experience tuples as discrete retrievable cases, improving from 78.65% to 84.47% accuracy over five iterations through memory accumulation alone.

The implication is structural. In the fine-tuning paradigm, the moat belonged to whoever had the most compute. With memory-augmented learning, the moat shifts to whoever has the best domain-specific experiences. A startup with deep expertise in legal contracts can build agents that outperform general-purpose models on contract analysis. They do this not by outspending frontier labs, but by accumulating superior execution data through its orchestration platform.

§

Execution Data

The millionth customer record adds less insight than the thousandth. The data grows. The value plateaus. Execution data behaves differently. Each interaction widens the gap. The result is winner-take-all dynamics within specific domains and a new form of capital the market does not yet know how to price.

This creates a category distinct from both enterprise data and training data. We call it execution data. Enterprise data answers “what happened”. Training data encodes “what patterns exist”. Execution data captures “what worked and why”. These are the specific sequences, strategies, and decisions that succeeded or failed in actual operation.

The economic properties are what matter. Traditional data shows diminishing returns. The millionth customer record adds less insight than the thousandth. The data grows. The value plateaus. Execution data compounds. Early experiences establish basic competency. Later experiences build on those foundations. The thousand-and-first claims processing discovers a pattern that improves performance across an entire category. Each interaction widens the gap.

The result is winner-take-all dynamics within specific domains. The first legal AI to handle a thousand contract negotiations will not just have more data. It will have a richer understanding of the problem space, including rare edge cases and creative solutions that competitors cannot acquire without processing a thousand contracts themselves.

In financial terms, execution data represents a new form of capital: memory capital. Like intellectual capital or brand capital, it is an intangible asset that generates future returns. Unlike traditional intangibles, memory capital can be precisely measured in terms of number of cases, performance improvement per case, or retrieval accuracy. It can then be directly deployed. Infrastructure commoditizes. Domain-specific execution data compounds. The market is beginning to price this shift, though without fully understanding it yet.

§

Context Defensibility

A competitor can match your features, your pricing, and even license your models. It cannot manufacture the memory of ten thousand orchestrated interactions. The new system begins as a stranger. That is the switching cost that protocols cannot dissolve.

Context advantages become durable moats only when defended against replication. Two factors determine whether they do.

Proprietary access creates the first defense. Physical-world operators possess inherent advantages because their context arises from irreplaceable operational reality. Tesla accumulates driving context through millions of vehicles in diverse conditions. Manufacturing facilities generate production context through actual operations. Logistics companies accumulate transportation context through real shipments. These companies build on contextual foundations that digital competitors cannot replicate because generating the context requires physical infrastructure.

The same applies in enterprise software through a different mechanism. The insurance company that has processed 50,000 claims through its orchestrated workflow has accumulated context that a new entrant cannot acquire without processing 50,000 claims. The context is a byproduct of operation. You cannot buy it. You cannot synthesize it. You earn it through sustained execution.

Switching costs create the second defense. When context accumulates in an orchestration platform, switching means abandoning everything the system learned. A competitor can match features, pricing, even license identical models. It cannot manufacture the memory of accumulated interaction. The new system begins as a stranger. The deeper the context, the higher the switching cost. That’s because the gap between what the incumbent knows and what the new entrant must learn widens with every interaction.

Protocols sharpen this dynamic. MCP and A2A standardize how agents connect to systems, dissolving integration friction. But protocols cannot transfer context. They make it easier for agents to access data across systems. They cannot replicate the behavioral understanding, relational mapping, and temporal patterns that emerged from years of orchestrated operation. Companies that relied on integration friction face exposure. Companies whose moat is the quality and uniqueness of their context find that protocols reinforce their advantage, making the platform more connectable while leaving the context inimitable.

§

The Investment Test

The standard against which we evaluate every company in the investment universe: does it sit at intent origin, accumulate context that compounds with every interaction, and generate that context from operations a competitor cannot replicate?

The Second Law reframes how we evaluate companies in the agentic economy. The question is no longer “how much data does this company have?” It is “what kind of context is this company accumulating, how fast is it compounding, and how defensible is it against replication?”

Three pillars work synergistically:

Memory architecture determines how context is stored, retrieved, and deployed. These are the episodic, semantic, and procedural layers that transform raw experience into operational intelligence.

Context depth determines richness across behavioral, transactional, relational, and temporal dimensions.

Defensibility determines whether the context can be matched through proprietary access, regulatory barriers, or the sheer volume of operational history required.

Memory systems without rich contextual content provide structure without substance. Deep context without architectural sophistication to organize and retrieve it proves unwieldy and underutilized. Defensible context that agents cannot leverage due to infrastructure limitations delivers strategic value slowly and incompletely. The durable positions require all three.

Companies where data compounds are building moats that deepen automatically. Every transaction, every user, every interaction adds to the advantage. The models become more accurate. The predictions become more reliable. The new customer benefits from everyone who came before, and everyone who came before benefits from the new customer.

The most defensible positions combine proximity to intent, deep compounding context, and proprietary access that prevents replication. The orchestrator that sits at intent origin, accumulates operational context with every interaction, and generates that context from irreplaceable operations holds a position that is nearly impossible to displace. This is the standard against which we evaluate every company in our investment universe.

§

The Second Law

The first two laws tell you where to stand and what to accumulate. They do not tell you what happens when the ground shifts.

The first two laws identify where to position and what to accumulate. But they describe conditions, not dynamics. The agentic economy is not static. Workflows change. Competitors adapt. New entrants arrive.

The orchestrator that maintains its advantage is the one whose coordination intelligence compounds faster than the environment shifts, whose learning improves with scale in ways that specialists cannot match.

That is the Third Law.

The first two laws describe what you own. The Third describes how you work and whether that matters.

§

The Workflow Moat

The coordinator sees what the specialists cannot. Not because it is smarter. Because it occupies a different position in the architecture, one that reveals the patterns between domains rather than the patterns within them.

Consider the insurance workflow once again.

The master agent has processed 50,000 claims over eighteen months. Its pricing agent has been invoked on all 50,000. Both have learned. Both are better than they were on day one. But they have learned different things, and the difference is structural.

The pricing agent has learned to price better. Its risk models have sharpened. Its calibration has improved. It handles edge cases with increasing precision. This is specialist learning: deep, narrow, and subject to diminishing returns. There are only so many ways to improve pricing. The capability curve flattens.

The master agent has learned something else. It has learned coordination. It knows which specialists work well together and which combinations produce errors. It has discovered that certain claim types need medical records enrichment before pricing, while others can proceed directly to adjudication. It has noticed that claims submitted in Q4 from manufacturing clients require a different decomposition than identical-looking claims submitted in Q2, because year-end inventory adjustments change the risk profile in ways the pricing agent alone cannot detect.

This is meta-expertise: the knowledge of how to make domain experts effective together. It is the accumulated workflow intelligence of the Orchestration Layer that now directs routing logic, exception handling, specialist sequencing, and failure recovery. Every interaction teaches the coordinator something the specialists never see. The specialists see their inputs and their outputs. The coordinator sees the entire workflow: what was tried, what failed, what succeeded, why the sequence mattered, and how the context affected the outcome.

A new competitor can license every specialist agent in the system. It cannot download the coordination intelligence that makes those specialists effective together. It must learn it from scratch; one orchestrated interaction at a time, while the incumbent continues to compound.

This is the Third Law: workflow intelligence is the accumulated operational choreography of how work gets coordinated. It creates switching costs that grow with every interaction.

The question is: How strong are those switching costs?

§

The Weakest Law

The Third Law is the most vulnerable of the three laws for defensibility. The first two require assets that cannot be manufactured: relationships and operational history. The third requires patterns that, under certain conditions, agents can learn autonomously.

The claim that coordination compounds while specialist capability flattens is intuitive. It is also, as a universal law, overstated.

The Google-DeepMind-MIT study that reported 81% relative improvement in structured financial reasoning also found the opposite pattern in other domains^{214Googe Research & MIT, Towards a Science of Scaling Agent Systems, December 2025}. Across all benchmarks, the mean multi-agent improvement was negative 3.5%^{215Yubin Kim et al. (Google Research, Google DeepMind, MIT), "Towards a Science of Scaling Agent Systems," arXiv:2512.08296, December 9, 2025}. In sequential planning tasks, every multi-agent architecture tested degraded performance by 39-70%.

Coordination did not compound. It destroyed value.

The study also found that coordination benefits are task-contingent. They are dependent on whether the problem is decomposable and parallelizable, and not universal. It identified a capability saturation threshold: once single-agent baselines exceed approximately 45% accuracy, adding coordination yields diminishing or negative returns.

More troubling for the Third Law is what happens when specialists learn to coordinate themselves.

In December 2025, Meta FAIR published Self-play SWE-RL^{216Yuxiang Wei et al. (Meta FAIR), "Toward Training Superintelligent Software Agents through Self-Play SWE-RL," arXiv:2512.18552, December 21, 2025}, a study in which a single software agent with no coordinator and no multi-agent orchestration autonomously learned to navigate complex codebases through self-play. The agent generated its own bugs, solved them, and generated harder bugs based on its failures. It required only access to the raw codebase: no human-curated issues, no test suites, no workflow documentation. Through sustained interaction with the environment alone, the agent outperformed the human-data baseline across the entire training trajectory, improving 10.4 points on SWE-bench Verified.

The implications for the Third Law are inescapable. If an agent can learn the workflow logic of a complex software repository through observation and interaction, then what prevents a sufficiently capable agent from learning the workflow logic of a claims processing system? Or an integration layer? Or a supply chain?

The answer is not “nothing”. Real barriers exist. But the barriers are weaker than those protecting the first two laws. Intent capture (the First Law) requires a relationship with the human who expressed the goal. You cannot learn that from observation. Operational context (the Second Law) requires years of accumulated experience. You cannot synthesize 50,000 claims outcomes from scratch. Workflow choreography (the Third Law) requires understanding of how processes connect. That, as Self-play SWE-RL demonstrates, is increasingly learnable.

The Third Law is real. It is also the law most likely to be tested by the market.

§

When Workflow Holds

An integration middleware platform’s moat is choreography. That is precisely what protocols dissolve. A vertical insurance platform’s moat looks like workflow. It is actually context wearing workflow’s clothing. The distinction determines who survives.

If the Third Law does not hold universally, the investment question becomes: under what conditions does it hold? Three conditions determine whether workflow intelligence creates durable advantage:

When workflow embeds operational context that emerged from practice, not documentation. The claims processing logic encoded in an insurance platform is not merely a sequence of steps. It includes exception handling paths that emerged from thousands of edge cases, regulatory compliance requirements that proved necessary through actual enforcement actions, specialist routing patterns that were refined by observing which combinations produced errors. This is workflow intelligence that cannot be learned from a manual because it was never written in one. It emerged from sustained operational interaction. It is, in effect, the Second Law expressing itself through workflow.

When switching costs are behavioral rather than technical. Traditional software lock-in is technical: data migration costs, integration reconfiguration, and user retraining. Orchestration lock-in is behavioral. When an organization’s operational rhythm has adapted to how one orchestrator decomposes intent, validates outputs, and handles exceptions, switching means rewiring the operational patterns of every process that adapted to the incumbent’s coordination intelligence. The new system does not just lack data. It lacks the judgment that emerged from years of orchestrated experience. It begins as a stranger who is technically capable but operationally naive.

When the domain involves decomposable, multi-step workflows rather than sequential reasoning. Recall that research has demonstrated centralized coordination produced dramatic gains on parallelizable tasks such as financial reasoning involving multiple data sources, cross-reference synthesis, and distributed analysis. However, it destroys value on sequential constraint-satisfaction tasks where coordination overhead fragmented reasoning capacity. The Third Law holds in insurance claims, supply chain logistics, enterprise procurement, and regulatory compliance. It fails in domains where a single capable agent outperforms any coordinated team.

These conditions produce a diagnostic that maps directly to investment decisions. Consider an integration middleware platform. This company’s entire value proposition is connecting System A to System B with the correct data transformations and sequencing. Its moat is integration choreography. That is precisely what MCP and A2A dissolve.

The platform coordinates workflows but accumulates no proprietary operational context. Its moat is integration friction. When protocols standardize connectivity, the workflow knowledge becomes replicable. An agent using MCP can discover and connect to the same endpoints. The choreography the platform encoded over years becomes learnable in weeks.

This is the most exposed position in the agentic economy: workflow orchestration without context.

Now consider a vertical insurance platform that has amassed decades of encoded claims processing logic, underwriting rules, and policy administration choreography. This looks like a deep workflow moat. But examine what is defensible. It is not the workflow sequence. It is the domain knowledge embedded in the workflow: the understanding of how insurance works in practice, the exception paths that emerged from millions of claims, the regulatory requirements that were discovered through enforcement rather than documentation.

Strip away the context and the workflow is learnable. The moat is context (the Second Law) expressing itself through workflow (the Third Law). The position holds due to the Second Law’s advantage, not because of the Third.

The same pattern appears outside software. In February 2026, Goldman Sachs revealed it had spent six months embedding Anthropic engineers within its teams to build Claude-powered agents for KYC compliance and trade accounting. These are document-heavy, rule-intensive workflows that combine data extraction with regulatory judgment^{217CNBC, "Goldman Sachs taps Anthropic’s Claude to automate accounting, compliance roles," February 6, 2026}. The bank’s CIO described the agents as “digital co-workers” for processes that are “scaled, complex, and very process intensive”. The agents review documents, extract entities, assess ownership structures, and trigger compliance checks. Internal tests showed that onboarding timelines collapsed by approximately 30%.

This is a direct test of the Third Law. KYC compliance is a multi-step workflow involving identity verification, sanctions screening, beneficial ownership analysis, and ongoing monitoring. A third-party compliance platform that merely choreographs these steps by connecting document scanners to screening databases to case management systems holds a workflow moat that agents can learn. Goldman’s agents are learning it now.

However, the bank’s decades of compliance outcomes represent an invaluable advantage because they are an asset that the agent platform cannot replicate:

Which edge cases triggered regulatory action.

Which documentation patterns correlated with genuine risk versus administrative friction.

Which client structures required enhanced due diligence not because the rules said so, but because enforcement history revealed it.

That operational context was accumulated through millions of onboardings across every jurisdiction where the bank operates. It is what makes Goldman’s own orchestration position defensible, even as third-party KYC platforms face exposure. The workflow is learnable. But this kind of institutional memory is not.

The pattern for the Third Law then requires an important nuance: workflow moats are defensible to the extent they embed operational context. The choreography, the routing logic, and the integration patterns, these elements of pure workflow are learnable and therefore vulnerable. Workflow that encodes irreplaceable operational intelligence is defensible because of the context it contains, not because of the workflow itself. That makes the Third Law a derivative moat.

§

The Switching Cost Asymmetry

Traditional lock-in is visible, static, and technical. You can calculate the migration cost on day one. Orchestration lock-in is invisible, dynamic, and behavioral. The actual cost only becomes apparent when you try to leave and discover that your entire operational rhythm depends on coordination patterns that exist nowhere else.

Even where coordination learning is moderate, switching costs are real and they compound over time. Traditional software lock-in is largely static: migration costs remain relatively constant, driven by data transfer, integration reconfiguration, and user retraining. These are significant, but they are visible and quantifiable.

Orchestration behaves differently. It is dynamic. Each interaction deepens the system’s embedded knowledge through every optimized workflow, learned exception, and validated specialist combination. The gap between what the incumbent system knows and what a replacement must reconstruct widens continuously.

Crucially, this lock-in is not immediately visible. Unlike traditional systems, its cost cannot be fully modeled upfront because it resides in accumulated context: decomposition logic, exception handling paths, and coordination patterns that only emerge through sustained operation. Organizations only grasp the magnitude when they attempt to switch and discover that their operational rhythm depends on capabilities that exist nowhere else. Orchestration lock-in is invisible until you try to switch.

Most importantly, the nature of the lock-in shifts. Traditional lock-in is technical. These Traditional systems bind through technical dependencies while leaving underlying processes intact. When an organization switches from Oracle to SAP, the underlying business processes.

Orchestration systems bind through behavior. Switching does not simply require moving data. It requires reconstructing the decision-making patterns and operational habits that have adapted to the incumbent’s coordination intelligence. The new system does not just lack information. It also lacks judgment.

This is the genuine contribution of the Third Law to the investment framework. Even in domains where the coordination learning curve is moderate rather than exponential, where agents can, in principle, learn the workflow, the switching cost still grows with operational history.

The question for investors must ask is not: “Does coordination compound forever?” Rather, what matters is: “Do switching costs compound fast enough to create a defensible position before the next generation of agents learns to replicate the workflow?”

§

The Contestability Window

The window during which workflow intelligence can be accumulated is finite. The companies that establish coordination intelligence first will compound their advantage with every interaction. Later entrants start from zero in a market where the gap may already be structural.

The Third Law creates urgency even though it is the weakest of the three conditions. That’s because unlike previous technology cycles, the Orchestration Layer is not reserved for technology companies. Orchestration is defined by owning the control plane where intent enters and coordination happens. That control plane can live anywhere. Workday can be the orchestrator for HCM workflows. But so can an insurance company for claims processing, a bank for credit analysis, a manufacturer for supply chain optimization.

For the first time, the strategic layer of the stack is contestable by companies that own domain expertise, customer relationships, and operational context and not just companies that write software. Physical-world operators possess inherent advantages precisely because their workflow intelligence arises from irreplaceable operational reality. The insurance company that builds its own Orchestration Layer accumulates coordination intelligence that no third-party platform can replicate because the platform does not process the claims, does not see the adjuster patterns, does not experience the seasonal variations, and does not learn the exception paths that emerge from actual operations.

The risk for these companies is not capability. It is speed.

Early orchestrators compound their advantage with every interaction. Later entrants start from zero in a market where the leading orchestrator has thousands of workflows of accumulated learning. The gap is not just difficult to close. In domains where the Third Law holds it may become structurally unclosable. This is why the Third Law matters despite its weakness. It is not the foundation of defensibility. It is the accelerant. Proximity to user intent tells you where to stand. Context tells you what to accumulate. Workflow tells you that the window for doing both is closing.

The framework applied to a high-stakes case in the market.

Palantir is a company with extraordinary architecture, execution, and a $307 billion market capitalization^{218Market data as of April 10, 2026} as of mid-April 2026.

The company’s execution is not in dispute. Q4 2025 revenue of $1.407 billion grew 70% year-over-year, accelerating for the tenth consecutive quarter. Full-year revenue reached $4.475 billion with GAAP net income of $1.625 billion for a 36% margin. U.S. commercial revenue hit $507 million in Q4, up 137%. Free cash flow guidance of $3.93 – 4.13 billion means the company will generate roughly its entire 2025 revenue in 2026 cash flow alone^{219Palantir Technologies, Q4 2025 Earnings Release (Form 8-K, Exhibit 99.1), February 2, 2026}. Revenue per employee of roughly $1 million is among the highest in the industry, and projected to be much higher in near future^{220SaaStr, Palantir’s CEO: “We Will Grow 10x … With Fewer Employees Than We Have Today”, 2025}.

As of mid-April 2026, the stock traded at approximately 67 times trailing revenue, 204 times trailing earnings, and 98 times forward earnings^{220SaaStr, Palantir’s CEO: “We Will Grow 10x … With Fewer Employees Than We Have Today”, 2025}.

The market has effectively priced Palantir as the orchestrator of the agentic economy, the company that will sit at the center of the stack, directing agent workflows across the enterprise.

Has it earned that position? Let’s test its position against the Three Laws of the Agentic Era.

§

The Diagnosis

Palantir occupies Rings 2 and 3 without Ring 1.

We start by understanding Palantir’s overall structural position using the Three Ring framework. Palantir builds no foundation models. It is explicitly model-agnostic, routing through OpenAI, Anthropic, and others. Its Harness is AIP, Foundry, and the Ontology layer, and its workflow-owning deployments across defense, intelligence, and enterprise together constitute the most vertically integrated Ring 2 + Ring 3 position in the public markets. Palantir is built outward-in: from workflow ownership, through Harness, down toward intelligence as an interchangeable input.

However, Palantir’s position in the three-ring architecture is more complex than most software companies. It is not Ring 2 alone, not merely a Harness that makes intelligence operational. In defense and government, Palantir is Ring 3. Foundry and Gotham are the systems of record: the applications where military operations are planned, intelligence is analyzed, supply chains are tracked, and mission-critical decisions are executed. The Ontology encodes the operational reality of the organizations it serves.

Critically, even within its Ring 3 deployments, Palantir lacks Ring 1: it provides no intelligence layer of its own. It depends on external model providers such as, Anthropic or OpenAI, for the cognitive capability the Ontology orchestrates. This means Palantir spans Ring 2 and Ring 3 simultaneously while remaining absent from Ring 1. It is a Harness built on operational context of extraordinary depth, but one whose intelligence and intent surface are both supplied by others.

The diagnosis that follows tests whether the depth of that position is sufficient to satisfy the Three Laws and whether 954 deployments (customers)^{221Palantir Technologies Inc., Form 8-K, Q4 2025 Earnings Release, February 2, 2026; Palantir Q4 2025 Earnings Call Transcript, The Motley Fool, February 2, 2026} of unmatched depth can defend against platforms with shallower context but exponentially broader reach.

§

On the Second Law – Context Depth

Palantir passes. As high as any company in the market.

The Ontology is the genuine article.

It is not a database. It is not a semantic layer. It is not a knowledge graph. It is a programmable operational model of the enterprise, a system that maps real-world business objects such as orders, aircraft, supply chains, patients to digital objects with defined properties, relationships, permissions, and executable actions.

This semantic layer creates shared business language across data sources. A kinetic layer enforces validation, approvals, and audit trails. A dynamic layer enables real-time execution with governance. The software category it occupies barely exists. It is closer to an operating system for enterprise operations than to any conventional data platform.

The context depth is genuine, proprietary, accumulating, and defensible. On the Second Law, Palantir passes with distinction.

§

On the Third Law – Workflow Intelligence

Palantir passes within its deployments. But the delivery model that creates the workflow intelligence is the exact structural position the agentic transition compresses.

The Ontology is an operational control layer. It does not merely store information for agents to query. It defines what agents are permitted to do. When an agent proposes “order more jet fuel”, the Ontology determines which supplier, what approval rules apply, what safety constraints bind, what cost limits govern, and what execution workflow follows. Agents do not bypass the Ontology. They operate through it. If the Ontology is fully deployed, every action in the enterprise flows through a layer that encodes the decision rights, governance requirements, and operational logic of the organization.

This is workflow intelligence in the precise sense of the Third Law: coordination patterns that improve with operational history and create compounding advantages. The system learns which approval paths produce faster outcomes, which supplier configurations minimize cost, which compliance sequences satisfy regulators. Within its 954 customers^{222Palantir Technologies, Annual Report (Form 10-K), fiscal year ended December 31, 2025, filed February 2026 — "The company served 954 customers."}, this is a genuine and deepening moat.

The vulnerability is not in the quality of the workflow intelligence. It is in the delivery model that creates it.

Palantir’s Forward Deployed Engineers are the human experts who embed within customer organizations to build the Ontology for each deployment. These teams are the mechanism through which this workflow intelligence is constructed. The FDE model produces roughly 75% boot camp conversion and 137% U.S. commercial growth^{223Palantir Q4 2025 Earnings Release, SEC 8-K, February 2, 2026}. It also means each new customer requires months of deep implementation engagement. The intent surface grows linearly, one customer at a time, through a human-intensive process that is structurally the opposite of how orchestration platforms scale.

And it is the exact intermediary position that the agentic transition compresses. The FDE exists in the gap between data complexity and AI deployment capability. This is the same gap that intermediaries occupy in insurance, wealth management, and logistics.

Agents close that gap. Claude Code built Cowork in around two weeks^{224Boris Cherny (head of Claude Code), via X, January 12, 2026, confirmed by Fortune: https://fortune.com/2026/01/24/anthropic-boris-cherny-claude-code-non-coders-software-engineers/} with four engineers. Palantir’s implementation timelines run 6-12 months^{225Palantir Technologies, FY2024 Annual Report (Form 10-K): "Our sales cycle often lasts six to nine months but can extend to a year or more."}. Boot camps have compressed initial deployment to five days. The AI FDE agent performs complete data engineering builds from natural language. These are real migrations. But the migration must be completed before agentic competitors offer lower-friction alternatives that capture the Orchestration Layer while treating the Ontology as an operational substrate beneath it that can be queried.

Management frames customer concentration as strategy: “density of client base over volume”. That choice also means Palantir’s reach grows linearly rather than geometrically. Each customer deepens the moat. None broadens the surface through which agents and users interact with the platform.

On the Third Law, Palantir passes within its deployments, but only on a foundation whose scalability the agentic transition directly challenges.

§

On the First Law – Proximity to Intent

Palantir fails. This is what prevents it from being the orchestrator despite possessing the other two characteristics.

The Ontology is not the operating system of the enterprise. The orchestrator is the entity that tells the operating system what to do.

In the agentic economy, the platform that captures the user’s original intent becomes the Orchestration Layer. The platform that receives instructions from the orchestrator, no matter how sophisticated, is the tool.

If the workflow requires operational data governed by Palantir’s Ontology, the orchestrator routes an agent through it and The Ontology governs what that agent is permitted to do, which supplier it can select, which approval path it must follow.

The Ontology constrains. It does not direct.

Palantir has 954 customers. No user opens Foundry to express a business goal. They open it because someone has already decided it is the right tool for a specific operational workflow. Intent capture increases the probability of controlling orchestration. The absence of it makes controlling orchestration structurally difficult. You cannot solve an intent proximity deficit with better technology. You solve it with distribution. Two competing architectural models are emerging.

In the intent-first architecture, championed by Microsoft and OpenAI, the user expresses a goal to an AI assistant. The assistant orchestrates agents and routes them through whatever operational systems the workflow requires, including, where deployed, the Ontology. The power center is the AI assistant layer.

In the operational ontology architecture, championed by Palantir, the Ontology itself is the control layer. Users and agents interact with the enterprise through the Ontology, which defines the objects, relationships, permissions, and execution workflows that govern what can happen. The power center is the operational data model.

Both architectures have merit. But the First Law tips the structural balance. The entity that captures intent first controls routing. The entity that controls routing determines which systems are invoked, in what order, and for what purpose.

The Ontology is indispensable within the workflow. But the orchestrator decides whether the workflow runs at all, and whether it runs through the Ontology or around it. Indispensable is not the same as orchestrator. An operating system that every application must use is extraordinarily valuable. It is still not the application the user opens to state their goal.

§

The Mispricing?

Now let’s return to Palantir’s stock, which as of Spring 2026 was priced as the orchestrator of the agentic economy. It traded at roughly 67 times trailing revenue, 204 times trailing earnings, and 98 times forward earnings [Figure 56].

Consider what the market pays for the companies that the structural analysis identifies as actually building the Orchestration Layer at Rings 1 + 2 per supra. Anthropic was valued at approximately 27 times annualized revenue^{226Bloomberg, Anthropic Attracts Investor Offers at an $800 Billion Valuation, April 14, 2026; TechCrunch, Anthropic shrugs off VC funding offers valuing it at $800B+, for now, April 15, 2026; Sacra, Anthropic revenue, valuation & funding, 2026}.

Now consider what the market pays for software companies that are actively building the Orchestration Layer at Ring 3. Salesforce – Microsoft – ServiceNow can all be considered Orchestration Layer candidates. Yet they are trading respectively at ~4x, ~9x and ~5.8x EV/NTM value as of April 10th, 2026^{227TIKR, Salesforce vs ServiceNow: Which Beaten-down SaaS Stock Has More Upside?, April 2026; TIKR, Palantir Stock Has Dropped 31% From Its Peak, April 17, 2026; Multiples.vc, Salesforce Valuation Multiples, April 1, 2026; Multiples.vc, Microsoft Valuation Multiples, February 24, 2026; Multiples.vc, ServiceNow Valuation Multiples, March 28, 2026; Salesforce Form 8-K, Q4 FY26 Earnings Release}. Palantir is trading ~7x above that at ~46x EV/NTM revenue^{228Palantir Q4 2025 Earnings Release, February 2, 2026; TIKR, Palantir Stock Has Dropped 31% From Its Peak, April 17, 2026; Multiples.vc, Palantir Valuation Multiples, April 6, 2026; GuruFocus, Palantir Technologies EV-to-Revenue, March 26, 2026}.

Palantir^{229Raphaëlle d’Ornano, "Does Palantir have a Durable Growth Moat? Making sense of the valuation of an AI champion through AGI lenses", Medium} occupies a genuinely rare position. It spans Ring 2 and Ring 3 simultaneously. It is a Harness built on operational context of extraordinary depth. No Ring 1 provider has this. Few Ring 3 candidates have built anything comparable. The market is not hallucinating. It is pricing something real.

But occupying Ring 3 is not the same as capturing Ring 3’s value. To capture the value of the outermost ring, to be the orchestrator, the captain, the entity must satisfy the Three Laws. The Ontology passes the Second and conditionally the Third. It fails the First.

The stock is priced for an entity that directs the work. The architecture describes an entity that governs the work once directed. The distance between those two positions, between constraining and directing, is measured in hundreds of billions of dollars. Palantir’s architecture will survive the agentic transition. Whether its valuation survives is a different question entirely.

The Ontology is the most valuable crew member on the ship. It is still crew. The stock trades at a price that assumes it is the captain.

§

If Palantir is a striking example, the sorting that the Three Laws produce will occur company by company, moat by moat, workflow by workflow.

This is the subject of Part IV.

The Three Laws describe the conditions under which a company can become the orchestrator and sustain durable advantage in the Agentic Era. But markets do not price conditions in the abstract. They price the structural positions those conditions give rise to. When proximity to intent, depth of context, and control over workflow converge within the same domain, they do not remain as separate attributes. They collapse into a single, identifiable role in the system that transforms human intent into executed outcomes. That role is not captured by any existing category. It is not a product definition, a sector classification, or a stage in the software stack as previously understood. It is a position that emerges from the architecture itself.

The purpose of this chapter is to describe that position and define it clearly. Because once it is visible, it can be systematically identified across the market, compared across industries, and evaluated as the primary locus of value creation and capture in the new regime.

§

The Species

The company that satisfies all Three Laws is not a sector. It is a structural position that is identifiable, scoreable, and investable across every industry classification.

What emerges from that shared position is best understood not as a sector, but as a species: a class of firm defined by its structural role in the orchestration of work.

A company that captures intent at the point of origin, that owns the deepest and most operationally relevant context, and that directs the workflows required to convert both into outcomes occupies a position that is categorically different from one that performs only a subset of those functions.

This manifesto calls that position AGNT.

AGNT is not a stock ticker. It is a structural archetype that can be identified, scored, and compared across industries. It cuts through traditional boundaries because those boundaries were built for a different regime, one where software was a tool and operations were separate from it. The market’s sorting mechanisms via sector-based indices, thematic baskets, and the Rule of 40 were calibrated for a regime in which software competed for IT budgets and physical businesses competed on operational efficiency. GICS codes describe where a company was. The Three Laws of Agentic Value describe where a company is going.

An enterprise CRM platform can be an AGNT. It captures intent because a hundred and fifty thousand companies express sales goals inside it every day. It holds context in the form of decades of customer interaction data, pipeline patterns, and conversion histories across millions of accounts. It compounds workflow intelligence with every sales cycle it orchestrates. Apply the Three Laws. The position is defensible.

A global insurance carrier can be an AGNT. It captures intent because the policyholder is its customer. It holds context in the form of actuarial loss data built over decades of underwriting real risk with real capital. No foundation model can synthesize this from training data. It was generated by operations that bear real-world consequences.

In the Agentic Era that separation collapses. A freight carrier and an enterprise software platform may have more in common, architecturally, than either has with the company sitting next to it in a sector index. The entity that receives intent, deploys intelligence, and coordinates execution becomes the locus of control, regardless of whether it is labeled as a software company, a financial institution, or an industrial operator.

The implication is that similarity is no longer determined by sector adjacency, but by architectural equivalence. Two companies in different industries may occupy the same structural position, while two companies in the same industry may be separated by it entirely.

The Three Laws provide the mechanism for making that distinction explicit. Apply them, and the species reveals itself.

The question is not what does this company sell. The question is where does this company sit in the “Orchestration Graph” that we define below.

§

Four Combatants

The contest for the Orchestration Layer is fought by four categories of combatants, each with structural advantages and structural vulnerabilities.

The first combatant: the AI laboratories ascending the stack.

OpenAI and Anthropic did not begin as orchestrators. They began as model providers selling intelligence at the bottom of the stack. But they understand that models are commoditizing. The intelligence layer alone cannot sustain the valuations the market has assigned, especially ahead of their planned IPOs.

So, the laboratories are building upward: OpenAI with Frontier, a control plane for deploying and managing agents from any provider across enterprise systems. Anthropic with Claude Code, Cowork, and vertical plugins that extend orchestration from developers to knowledge workers to entire business functions.

The entry strategy is what we call the “coding wedge”, documented in software engineering, now understood as a deliberate go-to-market: win developers first, then use the trust, habits, and organizational adoption built inside engineering teams to expand outward into the rest of the enterprise.

Three traditional software defenses are breaking under pressure. Integration moats erode when agents route dynamically across any system with a protocol connection. UI moats erode when the primary user is an agent executing through APIs, not a human navigating an interface. Capability moats erode when foundation models compose specialized functions on the fly.

The labs’ advantage is technological fluency and speed of iteration. Their vulnerability is the Second Law. They possess intelligence without institutional knowledge. They have no claims data, no routing history, no actuarial tables, no supply chain memory. They are climbing toward the Orchestration Layer from below, and every rung they reach is contested by incumbents who have occupied it for decades. The race is intelligence ascending against context defending. Chapter 13 traces this in detail.

The second combatant: the software incumbents pivoting from storage to action.

Salesforce, ServiceNow, Workday, SAP, and Oracle hold the canonical data on which enterprises operate. Ripping them out requires migrating decades of operational records, retraining thousands of users, rebuilding integrations with dozens of downstream systems, and accepting the risk of operational disruption in mission-critical processes.

Their position is not in question. Their role is.

In the pre-agentic world, the system of record was also the system of action. When a sales rep wanted to update a deal, she opened Salesforce. The application was the interface where intent met execution. In the agentic world, intent starts in a conversation. A user tells an agent: “Update the pipeline forecast for Q2”. The agent calls into the CRM’s API, reads the data, performs the analysis, and reports back without the user ever opening the application. The system of record is still essential. But it risks demotion from the system where work is directed to the system where data is stored. The database behind the orchestrator rather than the orchestrator itself.

The central strategic challenge for every enterprise software incumbent is to make the transition from system of record to system of action, the platform that receives intent, deploys agents, and coordinates workflows end to end. Salesforce is attempting it with Agentforce and the repositioning of Slack as the enterprise’s intent surface. ServiceNow is building agentic IT service management. Those that succeed become orchestrators. Those that do not will retain data gravity but lose control of the workflow and the pricing power that accompanies it.

Their advantage is context gravity. This is the irreplaceable institutional data that already lives inside their platforms. Their vulnerability is speed. The AI labs are building toward the same position from the opposite direction, and the orchestration position compounds with every interaction. The window narrows with each quarter. Chapter 14 traces the sorting.

The third combatant: the throughput node.

Not every software company must become the captain. Some of the most durable positions in the agentic economy will belong to companies that never aspire to the apex of the orchestration graph, but whose context is so dense that every orchestrator must route through them.

A throughput node does not capture intent. It does not control the workflow. It holds the context that makes the workflow correct. When that context is sufficiently deep and semantically rich, the orchestrator cannot complete the work without passing through the node. It doesn’t matter who occupies the captain’s seat. These are the tools that step into the orchestration loop rather than waiting beside it. They expose their context through open protocols, becoming surfaces through which agents move work rather than destinations where humans perform it. A tool that remains a destination risks bypass. A tool that becomes a node becomes structural.

The distinction between a throughput node and a convenience layer is semantic density across the production chain. A design platform whose component libraries, interaction models, and brand systems are inputs to every downstream workflow in product development is a throughput node. A project management tool whose task metadata can be replicated by the orchestrator’s own coordination logic is a convenience layer. The Second Law identifies the difference. The market has not yet learned to price it. Chapter 15 traces the distinction.

The fourth combatant: the domain operators building from the inside.

Insurance carriers, banks, logistics companies, manufacturers, and healthcare systems are the companies that own the physical assets, the regulatory relationships, and the operational context that satisfies the Second Law most deeply.

These companies never considered themselves technology companies. They are becoming orchestrators by necessity, because the agentic transition presents a binary choice: build your own Orchestration Layer or watch someone else build it on top of your operations and capture the margin.

A global logistics carrier deploying a master agent to coordinate fleet routing, capacity allocation, customs clearance, and last-mile delivery is not buying orchestration from a software vendor. It is building orchestration on top of its own operational context, using foundation models as reasoning engines.

The carrier’s competitive advantage was always its physical network and accumulated operational knowledge. In the pre-Agentic Era, that advantage was constrained by the throughput of human managers. In the Agentic Era, the same advantage is unleashed via coordinated by agents that operate at machine speed across the full scope of operations.

The same logic applies across every Capex-heavy, operationally complex industry:

A construction equipment manufacturer whose agents coordinate fleet scheduling, predictive maintenance, and supply chain management across dozens of job sites.

An insurer whose agents process claims through orchestrated workflows built on decades of actuarial data.

A bank whose agents orchestrate credit analysis, risk assessment, and portfolio management across a multi-trillion-dollar balance sheet.

In each case, the physical assets and operational history provide the context. The foundation models provide the cognition. The Orchestration Layer is built by the enterprise rather than purchased from a vendor because that is where the value concentrates. Their advantage is irreplaceable context generated by decades of operations.

Their vulnerability is that they must build fastest from the furthest starting point. These companies are not accustomed to building software platforms. They must learn or partner at a pace that their organizational cultures have never demanded. The risk for all four combatants is identical: time. The orchestration position compounds. Early movers accumulate coordination intelligence that later entrants must rebuild from scratch. However, the Third Law is the weakest of the three conditions: pure workflow choreography is increasingly learnable by capable agents. What compounds durably is workflow that embeds irreplaceable operational context. This is the Second Law expressing itself through coordination. In domains where all three laws hold, where a company captures intent, owns deep context, and controls decomposable, multi-step workflows, the gap may become structurally unclosable.

§

The Two Arenas

The contest for orchestration is fought on two fundamentally different battlefields.

There may be three combatants. But they are not all confronting each other directly. Instead, there are two competitive battlefronts.

Arena One encompasses workflows that already run through enterprise software platforms: CRM in Salesforce, IT service management in ServiceNow, human capital management in Workday, ERP in SAP, design in Figma, storage in Box, etc. These are the digitized, process-heavy workflows that defined the SaaS era.

In Arena One, three outcomes emerge for the companies inside it. Some will be displaced entirely. For instance, the capability providers that monetized a single function that agents can now compose from general-purpose intelligence. Some will be absorbed into the execution layer, surviving as API endpoints invoked by orchestrators, retaining revenue but losing pricing power and strategic control. Some will become orchestrators themselves, leveraging existing customer relationships, context gravity, and institutional trust to claim the captain’s chair while partnering with AI labs for the intelligence they direct.

The sorting in Arena One plays out between AI labs attacking from above, software incumbents defending from within, and throughput nodes anchoring themselves as mandatory bridges that every orchestrator must cross.

Arena Two encompasses workflows that were never fully digitized. These are the operational, physical-world processes where human cognition coordinated physical assets: supply chains, manufacturing lines, logistics networks, construction projects, and clinical operations.

These workflows generated vast operational data but were never mediated by a software platform in the way that CRM mediated sales. The coordination lived in managers’ heads, in phone calls, in spreadsheets, and in tribal knowledge.

In Arena Two, the battle is to see which enterprises become orchestrators and build their own Orchestration Layers before third parties get there first. The LLM is the engine, not the captain.

The intermediary reckoning plays out almost entirely in Arena Two. This is the structural inversion where physical asset owners ascend as cognitive middlemen fall. The freight broker that owned no trucks, the insurance broker that sat between carrier and client, and the wealth advisor whose value was synthesizing information that an agent can now synthesize faster.

Each occupied a position between the asset owner and the customer, monetizing cognitive labor that is now being automated. When agents close that gap, the value will not automatically transfer to the AI lab or to the software platform. It could also be captured by the asset owner if they leverage the operational context by creating an effective Orchestration Layer.

Both arenas tap the labor market. They do so for different reasons.

In Arena One, the expansion is a budget migration. Software companies that cross the reliability threshold stop selling seats from the IT budget and start selling outcomes from the labor budget. The workflows already exist in digital form. The TAM expands because the buyer is no longer comparing $150/seat/month against a competing SaaS vendor. They are comparing the cost of the orchestrated outcome against the fully loaded cost of the knowledge worker the outcome replaces. The same workflows, repriced against a surface five to ten times larger.

In Arena Two, the expansion is a digitization event. These workflows were never mediated by software at all. The coordination lived in managers’ heads, in phone calls, in spreadsheets. There is no existing IT spend to displace, only labor spend that was never addressable by technology until agents crossed the capability threshold. The TAM does not expand from a smaller budget to a larger one. It materializes for the first time.

Arena One reprices existing digital workflows against the labor market. Arena Two brings non-digital workflows into the orchestrated economy for the first time. Both are measured in trillions. Arena Two is potentially the larger opportunity because it starts from a base of zero digital penetration.

The entire value pool is greenfield.

This is why the agentic transition is not a story about software eating itself. It is a story about who captures the value when intelligence becomes abundant enough to coordinate both digital and physical operations at machine speed.

§

The Map

The remaining chapters of Part IV trace the sorting across each combatant and each arena.

Chapter 12 examines what the AI labs are building: not better models, but orchestration platforms competing for the enterprise control plane.

Chapter 13 traces the software reckoning: the sorting of enterprise incumbents into systems of action, throughput nodes, and casualties. The SaaSpocalypse was directionally correct and analytically lazy. The Three Laws identify which companies the market is punishing unfairly and which it is not punishing enough.

Chapter 14 examines how software tools defend: the throughput node strategy and the new network effects that emerge when agent density replaces user density as the driver of platform value.

Chapter 15 follows the leviathans: the physical-economy companies whose operational context, accumulated over decades, positions them as the most structurally advantaged orchestrators in the framework. The intermediary reckoning. The physical-asset inversion. The largest value migration that nobody is watching.

The companies that triggered the sorting face a sorting of their own.

§

The LLM battlefield has shifted. To understand the new frontline, however, one must understand how the terrain has transformed under the feet of the LLM giants since ChatGPT was first released in November 2022.

For much of the post-ChatGPT era, public attention remained fixed on t As 2025 began, agents were barely part of the discourse, let alone agentic AI. The benchmark to measure who was ahead in the generative AI race came down to model power. By the summer of 2025, conventional wisdom held the OpenAI had built a commanding lead in terms of model power, app downloads, compute deals, market share, and mindshare.

Less than one year later, conventional wisdom has been turned on its head. OpenAI is scrambling to catchup to Anthropic’s structural advantages. The models are commoditized table stakes, albeit extraordinarily expensive ones. And the LLM labs are leaving the model layer behind. They are building platform companies. The fight is now for the Orchestration Layer.

The distinction between these two layers is the difference between fragile and durable. Model intelligence answers the question: Can the AI do this task? Orchestration answers the question: Can the AI do this task, here, within this organization’s systems, rules, and workflows, reliably enough that the organization will depend on it?

The evolution of this competition was, of course, influenced by many of the technical and economic factors we have already defined. But the decisive shift that changed the trajectory of this battle and clarified the new agentic dynamic came from somewhere else, somewhere very specific. It came from coding.

§

The Coding Wedge

Software engineering is the only enterprise domain offering structured environments, objective ground truth, and high economic leverage in combination. That is why every lab’s orchestration strategy begins here, and why the coding wedge determines who controls the enterprise door.

Anthropic released an early version of Claude Code in February 2025^{230Anthropic, Claude 3.7 Sonnet and Claude Code, February 24, 2025} and then made it more generally available in May 2025^{231Anthropic, "Introducing Claude 4," May 22, 2025}. OpenAI unveiled Codex in May 2025^{232OpenAI, Introducing Codex, May 2025}. At first glance, these looked like developer tools. In fact, they marked a turning point in the enterprise adoption of agentic AI.The evidence was visible by mid-2025. In the “State of AI in the Enterprise” report by Menlo Ventures^{233Menlo Ventures, 2025 Mid-Year LLM Market Update: Foundation Model Landscape + Economics, July 31, 2025} released in July 2025, Anthropic’s share of enterprise budgets had climbed to 32% from 24% since the start of 2024, while OpenAI’s had declined from 34% to 25%. By the end of 2025^{234Menlo Ventures, "2025: The State of Generative AI in the Enterprise," December 2025}, Anthropic’s US market coding share had climbed further to 54% of the US coding market shared compared to 21% for OpenAI.

Again, the popular perception was that OpenAI was a juggernaut that was running away with the LLM prize. But in the trenches, it was Anthropic winning the battles for budgets and developers. As I analyzed this dynamic on behalf of clients to better understand what was driving this adoption, I gave it a name: The Coding Wedge.

Even now, months later, people view the impact of Claude Code with a facile, surface-level framing that is anchored in the previous paradigm. It’s a good product. Anthropic has strong marketing. It went viral. The brand resonated. Perhaps some or all of these are true at the margins. But these fail to recognize the core structural dynamic at work.

Software engineering was the first domain where autonomous agents crossed the threshold from experimentation to reliable production use. It is uniquely suited to this transition because it concentrates three properties that rarely coexist in other enterprise workflows:

Code lives in a highly structured environment. Every task requires sequencing, dependency management, decomposition, and recombination. Writing software trains models on the primitives of orchestration.

Code offers objective ground truth. It either compiles and runs, or it fails. That deterministic feedback creates a fast verification loop unavailable in most other domains. The model can act, test, correct, and improve against clear signals.

Code has unusually high economic leverage. Every productivity gain in software engineering compounds through the products that team builds, the workflows it enables, and the businesses those products support.

That combination made it the natural entry point for every lab seeking to move from model provider to orchestration platform. The adoption data reinforces this. Cursor, an AI-native code editor, raised funding at a $29.3 billion valuation^{235Cursor Blog, Passé, présent et futur (Series D announcement), November 13, 2025}. Claude Code surpassed $2.5 billion in ARR, with business subscriptions quadrupling since the start of 2026^{68Anthropic, Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation, February 12, 2026}. And 4% of all public commits on GitHub worldwide were authored by Claude Code^{236SemiAnalysis, Claude Code is the Inflection Point, February 5, 2026} as of February 2026, with as much as 20% projected by end of 2026^{235Cursor Blog, Passé, présent et futur (Series D announcement), November 13, 2025}. The coding wedge is not theoretical. Coding is the fastest-growing category in enterprise AI.

Rather than being about code completion, the coding wedge is about controlling the first enterprise domain in which agentic systems became reliably deployable. The deeper strategic point is even more important: coding agents are not merely coding tools. They are general-purpose agent Harnesses disguised as developer products.

That is what the coding wedge really is: the first opening through which orchestration enters the company. The progression was not accidental. It was strategic. Win developers through the coding wedge, then expand to the broader enterprise through the trust, behavioral patterns, and organizational adoption that coding establishes.

However, as noted in Chapter 2, Anthropic had another powerful tool in its arsenal: MCP. Released as open source in November 2024^{56Anthropic, "Introducing the Model Context Protocol," November 2024. Subsequent adoption data from npm download statistics and GitHub repository metrics.}, this protocol rapidly became the industry standard for agent-to-system integration, even adopted by OpenAI^{237OpenAI, Sam Altman on X, March 26, 2025} and Google^{238Google DeepMind, Demis Hassabis on X, April 9, 2025}. This gave Anthropic a structural edge. The axis of competition had rotated from “which model is best?” to “whose ecosystem is most deeply embedded?”.

That rotation could not be reversed by a single model release, however impressive. Anthropic’s coding advantage also became a powerful development lever. First, Claude Code won developers. That became a wedge to win enterprise customers. But it also became a method for learning the architectural strengths and weaknesses of agentic AI to extend it to general knowledge work.

§

Two Architectures, Two Bets

OpenAI is building a control plane that manages any agent. Anthropic is building an agent that becomes a platform. The divergence embodies different assumptions about whether intelligence commoditizes fully. The answer will determine hundreds of billions in enterprise value.

The new terms of this rivalry were on full display in early February 2026 as demonstrated by the volleys exchanged between the “Great Model Powers” across business systems. Every significant product that OpenAI and Anthropic have shipped in the final months of 2025 and the beginning of 2026 revealed that neither company believes the model layer alone can sustain their soaring valuations.

On February 5, OpenAI announced the latest release of Codex^{239OpenAI, "Introducing GPT-5.3-Codex," February 5, 2026}, its coding platform. This represented a critical moment for the company as it sought to regain ground in the critical fight for the coding wedge. OpenAI acknowledged this directly at the launch of Codex, describing coding as the foundation for a much broader class of knowledge-work agents. As part of that effort to extend its value to developers, OpenAI concomitantly launched Frontier^{240OpenAI, "Introducing OpenAI Frontier," February 5, 2026}, an enterprise platform for deploying and managing AI agents across business systems.

Meanwhile, Anthropic had leveraged Claude Code to create Cowork, and then the vertical plugins that triggered the “SaaSpocalypse”. And then a couple of weeks later closed a $30 billion funding round at a valuation of $380 billion68, more than double its $183 billion valuation from just five months earlier in September 2025^{241Anthropic, "Anthropic Raises Series F at $183B Post-Money Valuation," September 2, 2025}, and the second-largest private funding round in technology history. This all points to the real story.

Consider that $380 billion valuation. At $14 billion ARR, that represents a 27x multiple. If Anthropic is a model company selling intelligence in a market where intelligence is converging, then this is an extraordinarily aggressive bet on a commoditizing asset. But if Anthropic is a platform company, selling orchestration, workflow coordination, and accumulated enterprise context, then 27x may be the entry price for a generational enterprise franchise.

Claude Code, Cowork, the vertical plugins, Frontier, and Codex are not model improvements. All of them are orchestration infrastructure: systems designed to coordinate AI agents across enterprise workflows, accumulate institutional context, and create the switching costs that justify platform multiples.

And yet, the two leading AI labs have arrived at the Orchestration Layer with structurally different visions. Both have real products, launched within weeks of each other, that embody different assumptions about how the enterprise AI market will evolve. Understanding the divergence is essential for valuing the companies.

Anthropic pursued the same logic from the other direction: win developers first, then use the trust, habits, and infrastructure built inside engineering teams to expand outward into the rest of the enterprise.

Agents on Frontier receive employee-like identities with scoped permissions. They connect to data warehouses, CRM systems, and internal applications through what OpenAI calls a “semantic layer for the enterprise”. They build institutional memory from their interactions over time.

The most critical thing to know is that they do not need to be OpenAI’s agents. Frontier manages agents from Anthropic and Google. About 10 days later, on February 15, 2026, OpenAI staged a minor coup by announcing the hiring of Steinberger, the man behind OpenClaw, “to drive the next generation of personal agents”, according to CEO Sam Altman^{242OpenAI, Sam Altman on X, CNBC Coverage, February 15, 2026}.

§

Frontier is a Control Plane

OpenAI’s architecture positions Frontier as the management layer that sits above all agents, its own and everyone else’s.

Agents on Frontier receive employee-like identities with scoped permissions. They connect to data warehouses, CRM systems, and internal applications through what OpenAI calls a “semantic layer for the enterprise”. They build institutional memory from their interactions over time.

The design choices reveal strategic intent. Frontier’s agent identity management mirrors human HR systems: onboarding, scoped permissions, performance monitoring. OpenAI is not being metaphorical when it describes agents as “digital coworkers”. It is making an architectural claim. If AI agents operate as enterprise employees, the platform that manages them becomes as essential as the system that manages human employees. Frontier aspires to become the Workday for AI labor.

The most critical thing to know is that they do not need to be OpenAI’s agents. Frontier manages agents from Anthropic and Google. This appears to be generous. In fact, by welcoming competitors’ agents onto its platform, OpenAI concedes agent-layer competition in exchange for control-plane dominance. If enterprises standardize on Frontier for agent governance, OpenAI captures the value of orchestration regardless of which model powers any individual agent. The agents become interchangeable components. The control plane does not. This is horizontal platform logic. AWS applied to enterprise AI. It does not require OpenAI to have the best model. It requires OpenAI to have the best coordination infrastructure.

§

Cowork is an Agent that Became a Platform

Anthropic pursued the same logic from the other direction: win developers first, then use the trust, habits, and infrastructure built inside engineering teams to expand outward into the rest of the enterprise.

Rather than building a management layer on top of agents, Anthropic has built outward from the agent itself. The sequence is methodical and each step compounds on the last.

Claude Code proved the model could execute complex, multi-step enterprise tasks in the high-stakes domain of software engineering. It generated $2.5 billion in revenue68. MCP standardized how agents connect to external systems and was adopted by competitors, including OpenAI236 and Google237, establishing it as an emerging industry standard. Cowork extended orchestration beyond developers to knowledge workers. The vertical plugins created purpose-built entry points into specific business functions.

The strategic logic here is vertical integration. Anthropic is betting that agent quality remains meaningfully differentiated when embedded within orchestration. In that framing, the experience of using Claude to coordinate a complex legal review or financial analysis is sufficiently superior that enterprises will build their workflows around it. This is closer to Apple’s logic: control the end-to-end experience, build an ecosystem around your product, and make execution quality the moat.

But there is a subtle departure from the Apple analogy. MCP is open-source. Anyone can implement it. This appears to undermine lock-in.

The answer lies in a finding from enterprise AI adoption research: only 11%231 of enterprise builders switched AI providers, even though technical substitution was trivial. The lock-in is not technical. It is organizational. When an enterprise builds workflows around Claude’s specific orchestration patterns, meaning its approach to multi-step reasoning, its tool-use conventions, and its plugin interfaces, the cost of switching becomes operational.

You can swap the model in an afternoon. You cannot re-tune thousands of enterprise workflows in an afternoon. Anthropic is making a layered bet: MCP creates ecosystem breadth (every system connects), Claude’s quality creates ecosystem depth (Claude is preferred within that ecosystem), and organizational adoption creates inertia (no one switches even when they could). A market phenomenon of behavioral lock-in despite technical substitutability becomes a deliberate business strategy.

The two architectures represent different bets about the speed and completeness of model commoditization. If intelligence commoditizes fully, Frontier wins. The platform that coordinates interchangeable agents captures the greatest economic value. OpenAI wins not because GPT is superior, but because Frontier is the control plane. It wins even if Claude is the better agent, because the control plane sits above the agent layer.

If intelligence retains meaningful differentiation when embedded in orchestration, Cowork wins. Enterprises do not want “any agent, well-managed”. They want the best agent, deeply embedded. And the switching costs compound with every workflow built around it.

§

What $380 Billion Requires

Three conditions must hold for the orchestration thesis to justify the labs’ valuations. None is yet confirmed at scale.

If orchestration, not intelligence, is what drives the labs’ valuations, three conditions must hold:

Orchestration lock-in must prove durable, not merely behavioral. Enterprises are not switching AI providers today. But this was measured during a period of rapid growth when no one had reason to test the limits of their commitment. The real test arrives when Frontier offers agent-agnostic orchestration at compelling economics, or when a competitor undercuts on price. MCP is open-source. It creates connectivity but not captivity. For $380 billion68 to be justified, the network effects must become self-reinforcing: more workflows generate richer context, produce better orchestration, and attract more workflows. This flywheel is architecturally plausible. It is not yet empirically confirmed at scale.

The coding wedge must compound. Claude Code’s $2.5 billion ARR68 validates the entry strategy. But the path from coding to legal, finance, HR, and general operations is the critical progression. If the vertical plugins drive cross-functional adoption, the platform thesis holds. If coding remains the dominant revenue line while other verticals grow incrementally, Anthropic is a remarkable single-function business at enormous scale. That is a different valuation than an enterprise platform.

The labs must accumulate enterprise context faster than incumbents can defend. The systems of record hold decades of accumulated institutional context: the workflow knowledge, compliance history, and operational understanding that orchestration depends upon. The labs are starting from zero. Frontier builds “institutional memory” from agent interactions. That memory is weeks old. The institutional memory embedded in a Fortune 500 company’s Salesforce deployment spans years.

The labs have speed and adaptability. The incumbents have depth and irreplaceability. The labs must build enterprise context before model commoditization erodes the intelligence advantage that gives them the right to orchestrate. This is a race against the clock. The clock does not pause for fundraising announcements.

§

The Enterprise Tax

The market portrays AI labs as asset-light disruptors. The reality: to win the Orchestration Layer, they must build everything that makes enterprise software companies expensive.

There is a dimension of this transformation that the market has yet to price: the expense of becoming an enterprise software company. The market narrative portrays AI labs as asset-light technology companies that are disrupting bloated incumbents. The reality is more complex.

To win the Orchestration Layer, the labs must build what every enterprise platform company before them has built: field sales organizations, customer success infrastructure, compliance certifications, vertical domain expertise, and the organizational capacity to manage thousands of enterprise relationships simultaneously.

OpenAI has embedded Forward Deployed Engineers within customer organizations. Frontier’s enterprise customers require SOC 2 Type II, ISO 27001, and a suite of related certifications. The EU AI Act begins enforcement in August 2026^{243European Commission, Regulatory framework for AI (EU AI Act), 2024}, introducing compliance requirements that did not exist when these companies were founded. Enterprise sales cycles sometimes stretch twelve months or longer for major deployments. These demand specialized teams that change margin assumptions for the labs.

The $30 billion Anthropic raised in February 202668 was not exclusively for training runs. A meaningful portion of this capital will fund the construction of an enterprise go-to-market apparatus, including a sales force, compliance infrastructure, customer success organization, and vertical expertise. These are structural costs that permanently alter the business’s margin profile. A model company has research costs and compute costs. A platform company has all of these, plus go-to-market costs that scale with customer count rather than compute capacity.

This creates a tension that the labs’ financial disclosures will eventually make visible. The transition from model economics to platform economics requires near-term margin compression to build the switching costs and network effects that expand margins over time. The trajectory of that compression and subsequent expansion will tell investors more about the durability of these businesses than any benchmark or revenue growth rate.

Meanwhile, the labs face a clock that the incumbents do not. Model advantages erode quarterly as open source narrows the gap. If the labs do not establish orchestration lock-in before model commoditization is functionally complete, they risk becoming very expensive API utilities competing on price.

§

What This Means for the Public Markets

When the IPO prospectuses arrive, revenue from model API consumption and revenue from platform orchestration will appear identical in a financial statement. They are not identical in their implications for durability or terminal value.

When the IPO prospectuses for OpenAI^{244Wall Street Journal, "OpenAI Is in Talks With Banks About a Potential IPO," January 2026} and Anthropic^{245Reuters, "Anthropic considers IPO as soon as October 2026," March 27, 2026} arrive, they will present revenue growth, gross margins, and customer counts in the language familiar to technology investors. That language will be necessary but insufficient.

If this analysis is correct, if these companies are transitioning from model businesses to platform businesses, then the central valuation question is not the rate of growth but the nature of the growth. Revenue from model API consumption and revenue from platform orchestration may appear identical in a financial statement. They are not identical in their implications for margin trajectory, competitive durability, or terminal value.

The companies that complete this transition will justify platform multiples. Those that do not will eventually be priced as API providers in a commoditizing market, regardless of current growth.

§

The Single Points of Failure

Even if the orchestration thesis is correct and the labs successfully transition from model companies to platform companies, two structural vulnerabilities could undermine the entire strategy.

The first is open-source compression. For now, enterprises overwhelmingly prefer closed-source frontier models today. The MAP study^{246Melissa Z. Pan et al., "Measuring Agents in Production," arXiv:2512.04123, December 2, 2025 (v3: February 3, 2026)} found that 85% of production agent teams build entirely in-house using direct model API calls, foregoing third-party orchestration frameworks. Open-source adoption is limited to edge cases: high-volume workloads where inference costs are prohibitive, or regulated environments that prohibit sending data to external providers.

Enterprise teams default to the best-performing closed-source model available, and runtime costs are negligible compared to the human experts the agents augment. This is encouraging for the labs. But it is a snapshot, not a guarantee.

If open-source models such as DeepSeek and Qwen continue to advance at the rate they had at the end of 2025 and the start of 2026 and the performance gap closes sufficiently, that 85% preference could erode rapidly, because the underlying driver is pragmatic performance selection, not structural loyalty. Enterprises test the top models and pick the best one. If an open-source model becomes the best, they will pick it instead.

The second is the compute trap. This is the risk that Dario Amodei himself has articulated with remarkable candor. Asked on a podcast why Anthropic does not spend more aggressively on compute given its belief that a « country of geniuses in a data center” is imminent, Amodei was honest about the financial fragility of the model: « If my revenue is not $1 trillion, if it’s even $800 billion, there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt if I buy that much compute” ^{247Dario Amodei, interview with Dwarkesh Patel, "Dario Amodei — We are near the end of the exponential," February 13, 2026}.

By the same measure, if Anthropic underspends, it misses potentially massive growth opportunities. This is an unprecedented situation in enterprise technology. No company in the history of the software industry has operated at this level of capital intensity with this degree of revenue uncertainty over such a compressed timeline.

Anthropic’s annual burn rate in early 2026 was running at approximately $7-8 billion^{248Anthropic press release, Series G Funding, February 12, 2026; Sacra, Anthropic revenue, valuation & funding, 2026} (one-third of revenue, per company guidance disclosed in confidential financials reported by The Wall Street Journal and Fortune in November 2025). The $30 billion in fresh capital from the February 2026 Series G provides runway^{249Fortune, November 12, 2025; TechCrunch, November 4, 2025}. But the runway at these burn rates is measured in years, not decades — and the compute commitments required to maintain frontier model performance only increase. Anthropic has committed approximately $80 billion in cloud-infrastructure spending to Amazon, Google, and Microsoft through 2029, with an additional 3.5 GW of compute capacity secured through Broadcom and Google partnerships beginning in 2027.

The tension is structural. To win the Orchestration Layer, the labs need frontier-quality models. You cannot orchestrate enterprise workflows with a mediocre model. To maintain frontier models, they need massive and growing compute investments. To fund those investments, they need revenue growth to materialize on schedule.

If the orchestration transition takes longer than the compute commitments allow, if enterprise adoption moves at enterprise speed rather than startup speed, the financial model fractures. Intelligence is commoditizing. Context compounds. Compute burns. The race against all three will determine if the labs can move up the stack fast enough, and effectively enough, to capture enough of the enterprise to justify their precarious economic positions.

The market’s panic is directionally correct and analytically lazy.

§

Let’s briefly take a closer look at the SaaSpocalypse. Though we’ve referenced it several times, the details are relevant to understanding SaaS companies in the Agentic Era.

On January 30, Anthropic released plugins to extend Claude Cowork’s functionality by automating specialized tasks within an enterprise department. One of the eleven plugins was a legal plugin, a domain-specific extension that automates contract review, NDA triage, and compliance workflows for in-house legal teams. It is a modest feature drop: an open-source plugin on GitHub, built in a matter of weeks.

The market’s response was anything but modest. In 48 hours, $285 billion in market capitalization evaporated^{250Bloomberg, Goldman Sachs equity research, and Financial Times market analysis, January-February 2026.} from global software and services stocks. Thomson Reuters fell 14%. RELX dropped 16%. CS Disco sank 20%, and LegalZoom plummeted 16%^{251Market data from February 2 to February 6, 2026}. A Goldman Sachs basket of U.S. software stocks recorded its largest single-day decline^{252Goldman Sachs equity research, GSXUSOFT index performance, February 2026.} since the tariff-driven selloff of April 2025.

Two days later, Anthropic released Claude Opus 4.6^{253Anthropic, Claude Opus 4.6, February 2026}, its most advanced model to date, featuring a one-million-token context window, record-breaking agentic coding benchmarks, and a new “agent teams” capability allowing multiple AI agents to coordinate on complex projects in parallel.

Opus 4.6 launched with a direct PowerPoint integration, a pointed challenge to Microsoft’s Copilot. This arrived 72 hours after OpenAI launched a desktop application for its Codex AI coding system, designed to transform software development into something akin to managing a fleet of autonomous developers.

Over the span of five days, the cumulative effect of these announcements shocked the markets into recognizing that the AI labs are no longer just building larger frontier models and then selling APIs to software companies. They are moving up the food chain and coming after the software companies’ markets.

While the market’s reaction may have been directionally correct for some categories, it has been analytically lazy. If code was the wedge Anthropic used to open the enterprise door, plugins are how it is walking through and going further inside. Then came Cowork, extending agentic capabilities to non-developers.

What spooked the market was the realization that this path works, and that it works fast. Anthropic did not need years of enterprise sales cycles to threaten Thomson Reuters. It needed a plugin built over the course of weeks by its own coding agent. The undifferentiated selloff treated Thomson Reuters, ServiceNow, Salesforce, and LegalZoom as though they faced the same threat. They do not.

The SaaSpocalypse narrative conflates two very different questions. Collapsing them into a single panic trade has created severe mispricing and a genuine analytical blind spot. This is a confused, simplistic attempt to price a structural shift in where control, coordination, and value capture will live as autonomous agents become the primary actors inside enterprises. In doing so, it misses where the next generation of durable enterprise power will actually reside.

The real question is not whether AI replaces software, but which layer of the stack becomes the locus of intent, decision-making, and execution, and which layers are reduced to interchangeable utilities. That distinction separates the vulnerable from the transformable, the doomed from the leaders.

This is not the death of SaaS. It is the Software Sorting, a structural re-ranking of where control, coordination, and value will reside when autonomous agents become the primary actors inside enterprises. To navigate the sorting process, we can start by asking two fundamental questions:

Is your software a system of record or a workflow wrapper?

If you are a system of record, can you become the system of action, the layer that tells AI agents what to do, and not just the database they query?

The answers separate the companies facing existential risk from those facing a value capture negotiation. Confusing the two is the most expensive analytical error in today’s market.

§

The Real Casualties

Workflow wrappers and thin AI applications share a common vulnerability: if your product is “model + wrapper + workflow” and the model creator can ship the wrapper as a plugin, the moat was never structural. It was temporal.

The greatest carnage will fall on some of the venture-funded wave of AI application companies built between 2023 and 2025: legal AI startups, AI-powered analytics platforms, AI writing assistants, and AI sales tools. To name just a few of the categories. These companies raised hundreds of millions collectively on a thesis that there would be a durable margin between what a foundation model could do and what an enterprise needed done, and that this margin would support an independent application business. That now looks flawed.

What Anthropic demonstrated, and what OpenAI’s Codex desktop app confirmed, is that the foundation model providers have no intention of respecting that moat. A leading legal AI startup’s own co-founder identified OpenAI as its largest potential competitor, even though OpenAI had invested in the company. Despite raising $300 million and building a solid client base, the startup faces the prospect of Anthropic offering the same specialized interfaces through a plugin built in weeks. From a technical perspective, a legal AI startup built on OpenAI and Claude’s API has a limited data moat and minimal switching costs.

The second category is the public-market equivalent: workflow wrappers. These are established SaaS companies that bundle general-purpose functions such as writing, analyzing, reviewing, or routing with a user interface, some integrations, and a subscription model. They typically do not hold unique data or serve as the core of an enterprise’s operational systems. LegalZoom’s 20% drop reflected market concerns that it focuses on standard legal workflows without proprietary data edges that might not withstand AI replication.

The common vulnerability is identical: if your product is “model + wrapper + workflow” and the model creator can cut out the middle, the moat was never structural. It was temporal. You were arbitraging a gap that is now closing at the speed of a plugin deployment. Palantir CEO Alex Karp^{254Alex Karp, Palantir Technologies Q4 2025 earnings call, February 2026} said as much on an earnings call^{255Palantir Technologies, Q4 2025 earnings call transcript, February 2026}: AI is now so good at writing and managing enterprise software that many SaaS companies risk becoming irrelevant.

For workflow wrappers and thin AI apps alike, the Three Laws are clear. These companies lack proximity to intent: the user starts elsewhere and the wrapper merely executes. They lack context depth: holding no proprietary data that compounds over time. And they lack coordination intelligence: performing single functions without the cross-system orchestration that generates learning. All Three Laws expose the same structural absence. The position was never defensible.

§

The System of Record: Different Question, Different Stakes

Salesforce is not LegalZoom. ServiceNow is not CS Disco. Workday is not a legal AI startup that built a wrapper around Claude’s API.

These companies are systems of record. They hold the canonical data that enterprises operate on. Salesforce owns the customer graph^{256Salesforce company filings and product documentation. ServiceNow, Workday, SAP, and Oracle descriptions based on respective 10-K filings and product architectures} across sales, service, and marketing. ServiceNow owns the operational workflow layer via IT service management, HR service delivery, and security operations. Workday owns the employee lifecycle, including payroll, benefits, workforce planning, and financial management. SAP and Oracle own the ERP backbone across supply chain, manufacturing, procurement, and financial consolidation.

Ripping these out is not about switching to a cheaper plugin. It means migrating decades of operational data, retraining thousands of users, rebuilding integrations with dozens of downstream systems, and accepting the risk of operational disruption in mission-critical processes. Nobody is building a homegrown CRM in Replit to replace their Salesforce instance. The market, in its panic, forgot this.

But systems of record face a different threat, one that is more subtle, and more consequential for long-term value capture. The question is not whether the system of record survives. It is whether the system of record captures value like a platform or gets priced like a database. This is the distinction between being the system of action and being the system of storage.

In the pre-agentic world, the system of record was also the system of action. When a sales rep wanted to update a deal, they opened Salesforce. When an IT manager needed to route a ticket, they opened ServiceNow. The application was the interface where intent met execution. That is where the user lived, and that is where the value accrued.

In the agentic world, intent no longer starts in the application. It starts in a conversation. A user tells an AI agent: “Update the pipeline forecast for Q2”. The agent calls into Salesforce’s API, reads the data, performs the analysis, and reports back. This all happens without the user ever opening Salesforce’s interface. Salesforce is still essential. But it risks being relegated from the system of action to the system of storage, just the database behind the LLM.

Systems of record are not dying. But their economic role in the stack is being renegotiated. The outcome determines whether these companies trade at approximately 3x revenue^{257Multiples.VC website. Average NTM Revenue multiple across listed US SaaS companies, March 2026} or 15x revenue. That’s a difference of hundreds of billions in market capitalization.

This brings us to the question the market should be asking: not whether the systems of record survive, but whether they can make the transition to being systems of action.

§

From Storage to Action: The Transition

Enterprise systems are shifting from passive storage to active orchestration. The companies that transform their systems of record into systems of action will control what agents do, not merely the data they access.

Right now, we early in the agentic transition, and the interface between AI agents and enterprise software is structured and mediated. Agents interact with systems of record primarily through APIs and protocols. Agentforce, ServiceNow’s workflow APIs, MCP server integrations. These are structured endpoints that expose enterprise data and functionality to AI agents in controlled, permissioned ways.

In this phase, the system of record retains significant structural leverage. It controls which data the agent can access, which actions the agent can take, and which workflows the agent can trigger. The API is a gate, and the incumbent holds the key. This is why Salesforce’s integration of MCP into Agentforce matters: it positions Salesforce as the mediation layer between AI agents and customer data.

The early evidence is encouraging for incumbents who are moving. Salesforce is positioning Slack as the intent gateway, the conversational surface where employees express goals that get routed through Salesforce’s Orchestration Layer. The Agentforce Command Center provides a unified view of agent activities; persistent memory ensures agents retain context across interactions. ServiceNow is making the analogous move with its workflow engine.

Both companies are attempting to become the Orchestration Layer, the system of action, rather than passively waiting to be queried. But this API-mediated phase is transitional. It will not last.

A new category of Large Action Models is emerging that points toward what comes next. Through these models, agents operate enterprise software through the user interface itself exactly as a human would. They reason, click, type, and navigate. They are pretrained to operate specific classes of enterprise tools such as ERP systems, CRM platforms, revenue cycle management, and IT service management.

The emergence of UI agents has two profound implications for the system-of-record question.

The first is that the API gateway, the structural leverage that incumbents currently enjoy, will erode. When an AI agent can operate any application by simply using its interface, it no longer needs the incumbent’s permission or the incumbent’s MCP integration. It just opens the application and works. This accelerates the urgency for systems of record to complete the transition to systems of action, because the API-era leverage they currently enjoy is a wasting asset.

The second implication is more consequential, and it cuts in the opposite direction. A UI agent, no matter how sophisticated, needs to know what to do, not just how to click. The agent needs to know which approval path to take for this specific deal, which escalation logic applies to this class of incident, and which exceptions are routine and which require human review. That orchestration logic is the accumulated context of how an enterprise actually operates.

This is precisely what a system of action provides. And it is where the transition from storage to action becomes the decisive question.

A system of record that remains passive storage, a database the UI agent queries for a number, will indeed be devalued. But a system of record that has become a system of action, that sits close to intent origin, that owns the orchestration logic, the decision trees, and the contextual understanding of this company’s specific processes, becomes the brain that tells the UI agent what to do.

The agent is the hands. The system of action is the brain. In the UI agent era, the interface is commoditized. The model is commoditized. The only durable moat is the orchestration context that makes the agent’s actions correct rather than merely fast. This is, stated differently, the argument of the Second Law: context is the moat. What the software sorting reveals is that law operating in real time across hundreds of billions of dollars in market capitalization.

§

The LLM Lock-In Paradox

The market’s panic assumes the LLM is the durable control point. It is not. The model layer is architecturally the most substitutable component in the entire stack and the labs know it.

The market’s panic in February 2026 was predicated on a simple theory: LLMs are eating SaaS, therefore LLM providers win and SaaS incumbents lose. But recall that Menlo Ventures’ data revealed that despite model performance convergence and unprecedented ease of technical substitution, only 11% of enterprise builders switched AI providers in the prior year. Orchestration lock-in was being caused by the organizational complexity of recalibrating prompts, validating tool interactions, and ensuring behavioral consistency.

That means that whoever gets there first will, by virtue of getting their first, hold a natural advantage. Now ask yourself: which lock-in would an enterprise CFO prefer: A CRM that holds fifteen years of customer data, process customizations, and institutional context that would take two years and $50 million to migrate? Or a foundation model that could be swapped for a competitor by changing an API endpoint?

CRM lock-in is deep, structural, and data-driven. LLM lock-in is behavioral and organizational. Real, but fundamentally more fragile. An enterprise can retune prompts in weeks. It cannot rebuild its customer data or HR graph in that timeframe.

§

What the Market Should Actually Be Pricing

The Agentic Era is a repricing event, not an extinction event. But so far, the market is repricing the wrong companies in the wrong proportions.

Enterprise software is not disappearing. It is being reorganized according to a different logic: not which company owns the data, but which company directs the work. For two decades, the dominant architecture of enterprise software was the system of record. Platforms competed to become the authoritative repository of enterprise data. Once installed, the database became the gravitational center of the workflow. Applications surrounded it, users logged into it, and value accumulated around the system that stored the truth.

The Agentic Era alters that architecture. Data remains essential, but storage alone no longer determines control. What matters now is which system receives intent, interprets context, and coordinates execution across agents, tools, and workflows.

Storage becomes a component. Orchestration becomes the strategic position. This is the transition from systems of record to systems of action.

The companies that succeed in this transition will not merely store enterprise context. They will operationalize it. Their platforms will receive goals, decompose them into tasks, deploy agents to perform those tasks, and integrate the results back into the enterprise workflow. They will sit between the human who wants something done and the agents that do it.

The companies that fail to make this transition will still exist. But their role will be different. Their systems will become data repositories queried by someone else’s orchestrator. The database will remain valuable, but the control plane will sit elsewhere.

The Three Laws provide the framework for understanding which outcome occurs. What the market is witnessing is not a collapse of enterprise software, but a re-ranking of it. Some platforms will ascend to become orchestrators of work in the agentic economy. Others will remain essential infrastructure. And some will be reduced to interchangeable execution components inside someone else’s workflow.

The distinction between those outcomes is not sector classification, brand strength, or short-term technological fashion. It is architectural position. Systems of record are not doomed incumbents. They are context-rich platforms facing a narrow window to reassert themselves as systems of action.

The Agentic Era will also create a secondary sorting between investors who understand how to make this distinction, and those who do not.

The tools that plug into the orchestration loop will not merely survive. They will compound.

Not every company in the agentic economy must become the orchestrator. Some of the most durable positions will belong to companies that occupy a different architectural role: the throughput node.

A throughput node is a tool that agents cannot route around. It does not capture intent, and it does not control the workflow. Instead, it holds the context that makes the workflow correct. When that context becomes sufficiently deep, every orchestrator, no matter who occupies the captain’s seat, must pass through the node to complete the work.

In the orchestration graph, these companies may not be the captain. But they are an indispensable bridge.

§

Stepping Into the Loop: The Architectural Move

In the pre-Agentic Era, software tools were destinations. In the Agentic Era, the ones that survive become surfaces through which agents move work. These are bidirectional bridges that agents must pass through, not endpoints they can route around.

The strategic move for a software tool in the Agentic Era is simple: step into the orchestration loop. Historically, most software tools lived outside the loop. Humans moved information between systems manually: exporting files, copying links, or translating work from one tool to another. In the agentic workflow, those boundaries dissolve. Agents move work between tools directly through protocols.

This creates a new architectural question for every software company: will agents route through your platform, or around it? Tools that expose their context through open protocols position themselves as nodes within the orchestration graph. Instead of being a destination where humans perform work, they become surfaces through which agents move work.

The distinction is subtle but profound. A tool that remains a destination can be bypassed, while one that becomes a node turns structural.

§

The Figma Example

Design platforms make this dynamic visible.

Modern design systems extend well beyond visual artifacts. They encode component libraries, design tokens, layout hierarchies, interaction models, and brand systems that govern how software interfaces are built. This is semantic context, not decoration. Without it, code generation produces interfaces that function but fail to match the system the company ships.

Figma’s integration in February and March 2026^{258Figma Blog, "From Claude Code to Figma: Turning Production Code into Editable Figma Designs", February 17, 2026 ; Figma Blog, "Agents, Meet the Figma Canvas", March 2026} with AI coding agents demonstrates the architectural move. In June 2025^{259Figma Blog, "Introducing our Dev Mode MCP server", June 2025} the company released a development server that allowed coding agents to pull structured design context directly from design files. Components, styles, variables, and tokens could flow outward from the canvas into code generation workflows.

In February 2026^{260Figma Blog, "From Claude Code to Figma: Turning Production Code into Editable Figma Designs", February 17, 2026}, a partnership with Anthropic completed the loop. Developers working in Claude Code can push a rendered interface directly back into Figma’s canvas as fully editable design layers. This is not just a screenshot, but real design objects that designers and product teams can refine.

The workflow now runs both directions: Design context flows from the canvas to AI coding agents.

Generated interfaces flow back into the canvas for evaluation and refinement.

By building this bridge on an open protocol, Figma has done something strategically important. It has inserted itself directly into the orchestration graph. The platform is no longer merely where designs are created. It is where design context enters, and where agent-generated interfaces must return to maintain fidelity.

Agents can generate code anywhere. But when design integrity matters, the workflow routes through the design system. That is the definition of a throughput node.

The quarterly data confirms the dynamic is already operating. Weekly active users of the platform’s AI features grew over 70% quarter-over-quarter. More than half of enterprise customers above $100,000 ARR were building with it weekly, as of late 2025.

The most revealing number: nearly 60% of AI-generated files came from non-designers^{261Motley Fool, Figma (FIG) Q4 2025 earnings call transcript, February 18, 2026}. The surface is expanding beyond its original user base because the agentic workflow demands a collaborative evaluation surface, not just for designers, but for anyone shaping product decisions.

§

The New Network Effects: Agent-Density, Not User-Density

In the SaaS 2.0 paradigm, network effects accrued through user density. Protocol-based interconnection creates a fundamentally different dynamic: agent-to-tool-to-agent. Each new connection increases the routing density of the entire graph.

This reveals a new species of network effect that the market has not yet learned to price.

In the SaaS 2.0 paradigm, network effects accrued through user density. More designers meant more shared components, more reasons for the next designer to join. These effects were real but bounded. They scaled within a professional community and slowed as penetration matured.

Protocol-based interconnection creates a fundamentally different dynamic. Every tool that exposes its context through MCP becomes more valuable to every agent that can consume that context. Every agent that routes through protocol-connected tools creates a reason for more tools to join the graph. The network effect is not user-to-user. It is agent-to-tool-to-agent. Each new connection does not merely add a node. It increases the routing density of the entire graph.

The Third Law of workflow intelligence operates through this mechanism. The tool’s workflow intelligence deepens not because it gets better at design per se, but because it learns, through the density of agent interactions, which design patterns produce the best code, which component structures are most reliably interpreted by agents, which handoff workflows minimize error.

This is one of the domains where the Third Law holds: the workflow is decomposable, multi-step, and semantically dense. These are exactly the conditions under which workflow intelligence compounds rather than commoditizes.

§

Throughput Nodes vs. Convenience Layers

Semantic density across the production chain is what separates a tool agents must route through from a tool agents can route around.

Not every software tool can occupy this position. The difference lies in semantic density. Figma succeeds here because of a quality that not every horizontal tool shares: semantic density across the production chain.

Design touches product interfaces, marketing assets, presentations, websites, and developer handoffs. The canvas is a natural waypoint because the artefacts it produces are inputs to nearly every downstream workflow in product development.

When the majority of customers already use multiple products and the majority of AI-generated files come from non-designers, the company has crossed from design tool to cross-functional collaboration surface. That breadth creates routing density: enough surfaces touching enough workflows that agents encounter it as a natural node rather than an optional detour.

Not every horizontal SaaS company possesses this. A project management tool that coordinates tasks across teams performs a valuable function, but it risks reduction to a task-routing API if agents grow capable enough to handle coordination natively within the orchestrator. If we apply the Second Law, its task metadata is semantically thin. The distinction is between throughput nodes, where work must pass to maintain fidelity, and convenience layers, where the routing function can be absorbed by the orchestrator itself.

§

The Wrong Lenses

The undifferentiated selloff punished companies with strong throughput-node characteristics as severely as it punished convenience layers and wrappers. The Three Laws, applied company by company, reveal a dispersion in architectural resilience that sector-level analysis cannot detect.

The market still evaluates software primarily through the lens of the SaaS era: sector classification, category leadership, and user growth. The agentic transition introduces a different sorting mechanism. What matters now is architectural position:

Some platforms will become orchestrators.

Some will become throughput nodes.

Others will be absorbed into the execution layer.

The distinction does not follow sector lines. It follows the depth of context a system holds and whether agents must route through that context to complete the work. The SaaS playbook rewarded companies that built the best tool for a given task. The orchestration era rewards companies that become nodes the workflow cannot bypass. Tools that step into the orchestration loop will not merely survive the transition.They will compound through it.

§

Two durable categories have now emerged across enterprise software. Systems of record that evolve into systems of action, capturing intent and directing workflows. And throughput nodes, the tools with semantic density so deep that agents must route through them to preserve fidelity. Both outcomes belong to the software arena.

The largest category of winners may lie in companies the market has never considered technology plays at all.

In an economy of abundant cognition, operational context becomes the scarce asset. The largest value creation of the Agentic Era may not occur in companies building AI. It could occur in the companies that use it: the industrials, financials, insurers, healthcare systems, and consumer platforms that constitute the majority of the S&P 500 and have never been considered technology plays.

§

The first shock of the agentic transition appeared to be contained within software. Agents threatened SaaS products. Software companies sold to IT budgets. The disruption seemed destined to remain inside the technology sector. That assumption held for precisely as long as it took for someone to demonstrate an agent performing work for which human intermediaries charge. Within days of the initial software selloff in early February 2026, the tremors spread to insurance brokers, wealth managers, property advisors, and freight brokers^{262CNN Business, Why the ‘AI scare trade’ might not be done, February 16, 2026}.

The pattern was not random. Agents had crossed a boundary. They were no longer competing for the IT budget. They were competing for the labor budget. And the labor budget, in many service industries, particularly professional services, is the largest cost line on the income statement, often comprising 30-60% or more of total costs or revenue^{263IbisWorld, Industries with the Highest Labor Costs, 2026}.

For every dollar enterprises spend on software, approximately six are spent on services^{264Gartner Newsroom, Gartner Forecasts Worldwide IT Spending to Grow 10.8% in 2026, Totaling $6.15 Trillion, February 3, 2026}. Gartner projects worldwide software spending of $1.43 trillion in 2026^{265Gartner Newsroom, Worldwide IT Spending to Grow 10.8% in 2026, February 3, 2026}. The global professional-services market is estimated at ~$6.7 trillion in 2026 (management consulting, legal, accounting, engineering, marketing, HR, and advisory)^{266The Business Research Company, Professional Services Global Market Report 2025}.

Recall that research has suggested that programming skills are required in 82.5% of knowledge work occupations^{267Wang et al., How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations, arXiv:2510.22780, November 6, 2025}, and that existing agents can complete readily programmable tasks 88.3% faster at 90 to 96% lower cost than human professionals. This is not a narrow finding about code. It is a measurement of how much professional cognitive labor is structurally exposed across the entire economy. The number is very large.

The market read this as destruction. It was correct that something was being destroyed. It was wrong about what would replace it.

§

The Intermediary Architecture

Across insurance, freight, finance, real estate, and professional services lies a common structure: cognitive intermediation.

An insurance broker compares policies across carriers. A freight broker matches loads to capacity. A wealth advisor allocates portfolios. A property consultant negotiates leases. Different sectors. Same architecture. Each sits between a buyer and a seller and charges fees for cognitive work that involves comparing, synthesizing, optimizing, and coordinating.

For decades this architecture was economically sound. Information was scarce, computation was expensive, and expertise required years of training. But the cognitive tasks that justify intermediary fees decompose into two categories^{268Agrawal, Gans & Goldfarb, Prediction, Judgment, and Complexity: A Theory of Decision-Making and Artificial Intelligence, in The Economics of Artificial Intelligence: An Agenda, University of Chicago Press, 2019}:

• Intelligence: the application of rules, however complex, to structured problems. Comparing insurance quotes is intelligence. Calculating portfolio allocations is intelligence. Matching freight loads is intelligence.
• Judgement: decisions under genuine ambiguity where rules do not determine the answer. Advising a CEO on whether to self-insure a novel risk is judgement. Structuring a cross-border acquisition is judgement.

Most intermediary businesses bundle the two together and charge for both. The fee does not distinguish between the 80% that is intelligence and the 20% that is judgement. The client pays for the bundle.

Agents unbundle it. The intelligence work is precisely what production AI agents already perform.

What remains is a smaller set of activities: relationship management, exception handling, and genuine strategic judgement under irreducible uncertainty. These are valuable. They are also a fraction of the work for which intermediaries currently charge, and that is about to be repriced.

The question is not whether the intermediary is compressed. It is who captures the margin that is released.

§

What Happens to the Intermediaries

The intermediary does not uniformly die. The sorting is more specific.

Some intermediary functions disappear entirely. Where the work is pure intelligence, agents absorb it. Foundation models perform basic insurance comparison, straightforward tax preparation, and simple contract review without anyone building a specialized product. This is the low end of the intermediary market, high volume and thin margin, and it is being absorbed quietly by general-purpose AI.

Some intermediary functions are reinvented by AI-native services companies that sell the work, not the tool.

These companies bypass the intermediary and sell outcomes directly to the enterprise: drafting the NDA for the company rather than billing through outside counsel, handling the claim without the third-party adjuster, placing the insurance without the broker. The AI-native services company is a new and more efficient intermediary. It is real, fast-growing, and structurally advantaged in fragmented markets where no single asset owner controls the customer relationship.

Some intermediary functions are absorbed by asset owners reinternalizing. Where the asset owner’s context gives its agent a quality advantage over any outsider, and where the asset owner can reach the customer through AI-powered interfaces, the intermediary’s margin migrates back to the entity that generated the underlying value.

Some intermediary functions persist. Where the work requires genuine judgement under irreducible ambiguity, such as advising a CEO on a novel risk structure to take our previous example, or any act where the consequences of error are severe, skilled professionals augmented by agents will command premium fees.

The intermediary that survives is the one whose value was never cognitive arbitrage. It was judgement, trust, and institutional knowledge that resists extraction.

The sector bifurcates. The middle faces compression from every direction simultaneously. These are the large intermediary organizations built on the premise that scale in cognitive labor creates a moat. Above and below, the margin narrows. What remains is either reinvented as an AI-native service or concentrated in the hands of fewer, more specialized professionals operating at a higher level.

§

The Scarce Asset

For a generation, markets rewarded asset-light business models.

The freight broker without trucks traded at higher multiples than the carrier. The wealth platform that held no assets was valued more than the bank. The advisory firm commanding expertise was priced above the company owning the physical infrastructure. The logic was simple: in a knowledge economy, cognitive capital outperforms physical capital. Brains beat iron.

Agentic AI inverts this. If cognition becomes abundant, cognitive capital loses its scarcity premium. The functions such as analysis and coordination that once justified the intermediary’s existence, become available at commodity pricing from multiple sources. What remains scarce is operational context. The accumulated intelligence generated by actually running a business over time.

A global logistics carrier’s knowledge of lane pricing dynamics, seasonal patterns, equipment failure modes, and delivery performance across every corridor it serves. An insurer’s actuarial loss data built over decades of underwriting real risk with real capital. A bank’s century of credit decisions with outcomes observed. A pharmaceutical company’s clinical trial archive. A manufacturer’s performance data.

This context is not data in a warehouse. It is the institutional intelligence that emerges from sustained operation – from the interaction between millions of decisions, outcomes, and corrections accumulated over time. It tells the agent what to do next in this specific situation based on everything the organization has learned.

A competitor could, in principle, license the underlying data. An insurer could sell its loss tables. A carrier could export its historical shipment records. But what transfers in that transaction is a snapshot of what was true – a stock that begins depreciating the moment the operation stops feeding it. What does not transfer is the flow: the continuously generated understanding of how lane dynamics shift, how risk profiles evolve, how seasonal patterns interact with capacity constraints in ways that only sustained operation reveals. The data is the exhaust of operations. The context is what the memory system produces when it consolidates that exhaust over years. You can buy the exhaust. You cannot buy the engine.

Nor can you replicate it from public sources. You cannot simulate decades of clinical trial outcomes. You cannot synthesize an actuarial portfolio from training data. You cannot reconstruct the institutional understanding of a logistics network from its transaction logs, any more than you can reconstruct a surgeon’s judgment from the operating room’s medical records. The records document what happened. The judgment – the context – is the product of having been there, every day, for ten thousand cases.

The word “asset” requires expansion beyond the physical. A FedEx truck is an asset. So is Eli Lilly’s clinical database. So is JPMorgan’s credit decision history. So is Disney’s library of intellectual property and the audience data generated by streaming it. The Leviathan is defined not by physicality but by operational irreplaceability. This is context so deep that no new entrant, however well-funded, can replicate it without running the same business for the same duration.

§

The Knowledge Architecture

Possessing the asset and deploying it are not the same thing.

An agent cannot traverse institutional memory that exists in the heads of experienced operators, scattered across document archives, or buried in databases designed for human queries. The context described above is real. In most companies it is also inaccessible because it is locked behind architectures built for a different era.

Between accumulated intelligence and the intelligent agent lies a gap that is not operational or cultural but architectural. The architecture that closes it is the knowledge graph.

A knowledge graph encodes institutional intelligence as a network of typed entities connected by typed relationships, semantic triples of the form (entity, relationship, entity) that represent how things in the operation relate to one another^{269Belova et al., An Alternative Trajectory for Generative AI, arXiv:2603.14147, March 14, 2026}. Each triple is individually interpretable, auditable, and expandable. Together they form a machine-traversable model of how it works that is causal, relational, and temporal.

While a database stores what happened, a knowledge graph encodes what it means, what caused it, what follows from it, and what to do about it. The graph is a dynamic symbolic layer that grows with every operational event, every decision recorded, and every outcome observed. New nodes and edges are added as the business runs. The graph compounds.

Neurosymbolic AI is the formal integration of symbolic reasoning with neural networks. Recent work in this field has converged on a structural claim: knowledge encoded as explicit graphs of entities and relationships provides a “structural counterweight” to the hallucinations and factual mistakes observed in purely neural systems^{270Margarita Belova, Yuval Kansal, Yihao Liang, Jiaxin Xiao et Niraj K. Jha, "An Alternative Trajectory for Generative AI", arXiv, arXiv:2603.14147 [cs.AI], 14 March 2026}. The language model does not need more data. It needs data with structure in the form of semantic relationships it can traverse, verify against, and be constrained by during generation.

The implications for high-stakes domains are immediate and unforgiving.

In process industries where a wrong recommendation can be catastrophic, the knowledge graph is becoming a prerequisite. Think healthcare where a hallucinated diagnosis carries liability or financial services where regulatory bodies require auditability.

Frameworks such as GraphRAG (Microsoft-developed framework that enhances Retrieval-Augmented Generation (RAG) by extracting knowledge graphs from text data, organizing them into hierarchical communities, and using graph traversal for more accurate, context-rich retrieval)^{271Luo et al., Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models, ICML 2025 / arXiv:2410.13080, October 2024 (v2 May 28, 2025)} and Graph-Constrained Reasoning (framework that integrates knowledge graphs directly into LLM decoding via a KG-Trie structure, constraining reasoning paths to valid graph edges to eliminate hallucinations and ensure faithful, interpretable outputs on graph data)^{272Microsoft Research, GraphRAG: Unlocking LLM discovery on narrative private data, April 2, 2024} have demonstrated that graph-grounded generation achieves measurably higher factual accuracy than retrieval alone, with some architectures achieving formally zero hallucination on domain-specific benchmarks by constraining the model’s output to paths that exist in the graph.

The company whose agents reason over a knowledge graph produces recommendations with provenance. Every output traces back through verified relationships to documented operational history. The company whose agents reason over raw data produces outputs that are plausible but unprovable.

In domains where the difference between grounded and plausible is the difference between a correct intervention and a catastrophic one, this is not a feature comparison. It is a structural divide.

§

Belova et al. (2026)^{273Margarita Belova, Yuval Kansal, Yihao Liang, Jiaxin Xiao et Niraj K. Jha, "An Alternative Trajectory for Generative AI", arXiv, arXiv:2603.14147 [cs.AI], 14 March 2026} articulate why this architecture may generalize far beyond process industries.

Their central argument is that the current trajectory of scaling monolithic models collides with hard constraints in energy, data quality, and depth of reasoning. Instead, they propose alternatives built on “domain-specific superintelligence”. These are specialized models trained on high-quality domain data, paired with explicit symbolic abstractions and modular reasoning engines.

They envision a “society of DSS models” where orchestration agents route tasks to specialized back-ends operating within graph-structured knowledge. This reflects a broader shared recognition that intelligence without structure is unreliable, and that the structure required is precisely the institutional intelligence that Leviathans have spent decades accumulating.

The companies that encode domain-specific relationships, constraints, and causal logic as graphs gain a compounding advantage that deepens with use. Those that leave it unstructured remain in possession of an asset they cannot fully deploy.

The knowledge graph is the mechanism that converts the scarce asset into deployable intelligence. With it, the Leviathan can reinternalize, extend the meta-layer, and compound. Without it, the context remains latent, real but inert.

§

Reinternalization

The first consequence of the inversion is that margin comes home.

For decades, enterprises outsourced cognitive functions because building them internally was too expensive, too slow, or too specialized. Insurance carriers outsourced distribution to broker networks. Banks outsourced advisory to independent wealth managers. Manufacturers outsourced logistics coordination to freight brokers.

Each decision was rational when cognitive labor was scarce. Each becomes irrational when cognitive labor is abundant and cheap.

The carrier that deploys its own agents reinternalizes the distribution function. It sells directly to the customer through conversational interfaces and browser agents. The broker’s commission, typically 5 to 15% of premium (and higher in specialty commercial lines)^{274Insure.com, How much do insurance agents make, June 26, 2025; KBI Benefits, Insurance Broker Compensation, 2024}, becomes margin recaptured by the asset owner. For a global insurer writing $30 billion in annual premium, the arithmetic is immediate: $3 to $6 billion in gross commission^{275Insure.com, June 2025 for rate base} expense becomes contestable.

The advantage extends beyond the direct channel. On any comparison surface where a customer evaluates multiple carriers, the insurer whose API returns a dynamically underwritten, context-grounded quote wins more business than the insurer returning a static rate table.

Better grounding produces better responses on every surface, including surfaces the asset owner does not own. The same logic operates across every intermediated industry.

A bank deploying AI advisory agents reinternalizes wealth management. A logistics carrier deploying routing agents reinternalizes freight brokerage. A pharmaceutical company deploying regulatory intelligence agents reinternalizes functions outsourced to CROs.

In each case, the asset owner possessed the context all along. What it lacked was the cognitive capability to deploy that context at the point of customer interaction. Agents provide the missing capability. The margin comes home.

§

The Meta-Layer

Reinternalization captures existing value. The second consequence of the inversion creates new value. The Meta-Layer is revenue that arises when operational context is deployed as intelligence inside customer infrastructure – creating entirely new business lines that did not exist before the agentic transition.

In this case, the asset owner takes the operational intelligence generated by running its core business and deploys it as agents that operate inside its customers’ infrastructure. It does this not as data feeds, not as dashboards, but as autonomous systems that produce measurable outcomes the customer pays for.

The agents become live conduits to the asset owner’s operational network, continuously tethered to real-time data that exists only because the asset owner runs the infrastructure. Disconnect the agent from the live data feed and it degrades to a static lookup table. The meta-layer generalizes wherever an enterprise possesses operational context that cannot be replicated without running the operation that generated it.

An insurer deploys risk assessment agents into commercial clients’ operations, flagging supplier risks before they become claims, predicting incidents before they occur. The client pays for loss reduction. The insurer captures new revenue while reducing its own claims exposure. A bank deploys credit intelligence agents into corporate clients’ treasury workflows, predicting payment delays, optimizing working capital. A construction equipment manufacturer deploys coordination agents on customers’ job sites, optimizing equipment utilization in real time based on weather, ground conditions, and crew scheduling.

The manufacturer is no longer selling machines. It sells construction orchestration as a service. The hardware becomes nodes in an intelligent network. The Orchestration Layer becomes the margin. In each case the meta-layer exhibits economics the core business does not. Revenue is recurring rather than transactional. It compounds through the Third Law (defined in Chapter 10) as workflow intelligence deepens. It is additive, layered on top of existing assets, not substitutive.

The switching costs are operational. The cost of losing coordination intelligence built over years of embedded operation.

§

The AI Compounder

The market recognizes AI enablers: companies selling the infrastructure. It recognizes AI adopters: companies deploying agents to improve existing operations. Neither category captures what is forming inside the Leviathans.

The AI compounder is a company restructuring its operations to capture margin expansion through reinternalization and revenue transformation through the meta-layer, grounded in operational context that satisfies all Three Laws of value simultaneously, and with intelligence at its core.

The transformation will not manifest linearly. Companies must redesign workflows around what agents do well rather than forcing agents to mimic human processes. The work spans years. Near-term results will look messy. Operating margins may compress as restructuring costs hit the income statement, and the market will punish the uncertainty.

Then the numbers will break through. Revenue growing while headcount declines, margins expanding on a restructured cost base, new revenue lines appearing from operational intelligence deployed as commercial agents, market share growing without increasing sales capacity.

The companies that complete this transformation will not be valued as industrials or financials or logistics or pharmaceutical companies. They will be valued as platforms because the meta-layer exhibits platform economics: recurring revenue, compounding network effects through agent density, and switching costs that grow with every embedded interaction. The core business provides the foundation. The meta-layer provides the multiple.

§

The Largest Arena

Follow the tremors. The labs’ platform build-out triggered the SaaSpocalypse and sorted Arena One into casualties, systems of action, and throughput nodes. Then the blast radius jumped beyond software.

This chapter traced the structural inversion that made operational context the relatively scarce and therefore increasingly more valuable asset of the Agentic Era, replacing a position traditionally held by intelligence, software, and data at the heart of Arena Two.

That inversion produces two forces.

Reinternalization recaptures margin from intermediaries. The meta-layer creates entirely new revenue by deploying operational intelligence as commercial agents inside customer infrastructure.

Together they produce the AI compounder. This is the Leviathan that rises from the non-technology economy with context no startup can replicate, switching costs no platform can match, and an earning power transformation the market has not yet learned to see.

The sorting determines who survives and who ascends. Part V traces the financial anatomy of this transformation: what the new earning power looks like when it appears on an income statement that the market’s models were never built to read.

Every paradigm produces its own economics: a distinct cost structure, growth dynamic, and distribution of value. This one is no exception. Understanding it is foundational to valuing everything built on top of it.

§

The Economy That Agentic AI Produces

A paradigm is not a tool. It is the organizing logic through which a technology reshapes how work is performed, how value is created, and how economic activity is structured.

The steam engine was a technology. Industrialization was the paradigm. The microprocessor was a technology. The knowledge economy was the paradigm. Agentic AI is the technology. Machines becoming actors is the discontinuity. The Inference Economy is the regime it produces. Orchestration Economics is how to assess value and moat durability within it.

The Inference Economy is the regime in which workflows consume compute as their primary variable input. The unit of economic activity shifts from the labor hour to the orchestrated workflow.

The cost of production shifts from human compensation to inference compute. The binding constraint on output shifts from the number of trained professionals an organization can hire, retain, and coordinate, to the volume of orchestrated workflows it can design, deploy, and scale.

§

Workflows Imply Compute

In the knowledge economy, a workflow consumed human cognition. In the Inference Economy, the same workflow consumes compute and human cognition. The structural shift is not that agents replace humans. It is that the primary variable input to the workflow changes.

An insurance claim passed through a chain of specialists each consuming labor hours, each adding cost proportional to compensation.

In the Inference Economy, the same workflow consumes compute. An orchestrating agent decomposes the claim and delegates to specialists: extraction, medical coding, fraud detection, pricing, and compliance. Each specialist processes its task by consuming inference tokens. The orchestrator coordinates, validates, and delivers [Figure 66].

The structural shift is not that agents replace humans. It is that the primary variable input to the workflow changes. Human judgment remains at the level of intent, oversight, and exception handling. The programmatic cognitive work migrates from labor hours to inference tokens. The cost difference between the two regimes is large. Even if research suggests the potential for agents to drive massive cost reduction at the task level, that does not necessarily translate directly to workflow-level savings.

Real orchestrated workflows involve coordination overhead, error handling, retry loops, human review gates, and exception routing. The actual cost of an orchestrated workflow is the sum of inference consumed by all specialist agents, plus the coordination overhead of the orchestrating agent, plus the cost of human oversight at quality-critical junctures. The savings at the workflow level are substantial but meaningfully lower than the task-level headline suggests.

Every workflow implies compute. Every orchestrated task consumes tokens produced by the inference infrastructure that is the largest coordinated private investment in peacetime history. The ~ $650 billion flowing into AI infrastructure in 2026147 is building the factories. The Orchestration Ratio determines how much of the economy buys what those factories produce.

§

The Orchestration Ratio

A single variable governs the transition to the Inference Economy: how much cognition shifts from humans to machines. The Orchestration Ratio captures this shift by tracking where value, cost, and control move as workflows become increasingly autonomous.

The governing variable of this system is the Orchestration Ratio: the share of cognitive work in any workflow performed by inference compute rather than human labor.

At a ratio of zero, the workflow is entirely human. This is the knowledge economy as it existed for decades. At a ratio approaching one, the workflow is fully autonomous. The global economy is moving from left to right on this spectrum, at different speeds in different domains. The pace of movement determines the cost structure, the growth dynamics, and the distribution of value across every industry.

The ratio has a ceiling. It rises if the marginal cost of the next unit of inference remains below the marginal cost of equivalent human cognition for a given task. If that condition reverses, the ratio stabilizes or declines.

The Orchestration Ratio can be quantified at multiple levels of granularity. For any given workflow it is formally defined as

where cognitive effort may be measured in tokens processed, discrete task completions, or cost-weighted contribution.

At the workflow level, the ratio is computed through one or more of the following proxies:

the share of total workflow cost attributable to inference compute rather than human compensation;

the share of discrete cognitive steps executed autonomously;

the inverse of the frequency and duration of required human oversight or exception routing;

the share of end-to-end workflow latency handled without human intervention.

At the company or department level, the Orchestration Ratio is the cost-weighted average across the organization’s portfolio of workflows.

At the sector level, the ratio aggregates the share of cognitive work that has migrated from labor hours to inference tokens across industries or the broader economy.

The ratio naturally reflects the distinction between commodity and frontier inference introduced infra. Workflows dominated by commodity inference exhibit faster gains in the ratio per dollar spent; frontier-heavy workflows show slower but higher-quality gains. The routing intelligence of the orchestrator therefore influences both the level and the slope of its Orchestration Ratio.

The Inference Economy is not an inevitability. It is a regime that holds if its foundational cost advantage holds. The structural evidence of inference costs declining at 5 to 10x per year^{276a16z, Welcome to LLMflation, November 12, 2024; Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, March 12, 2025} across multiple independent and reinforcing drivers suggests the advantage is robust. But that is a condition, not a law of nature.

NVIDIA’s Huang described the supply side of this regime at GTC 2026187: the data center as an AI factory whose product is the token, whose efficiency is measured in tokens produced per watt. The Orchestration Ratio describes the demand side: the share of economic activity that consumes those tokens, governed by the mix of human and agent cognitive input across the economy’s workflows.

Huang’s metric determines the floor cost of producing intelligence. The Orchestration Ratio determines how much of the economy consumes it. The Inference Economy is the product of the two: tokens produced at declining cost and consumed by an expanding share of economic activity.

§

Deflationary Growth

The Inference Economy has a macroeconomic signature that distinguishes it from every prior regime.

Deflationary growth refers to sustained economic expansion in which production costs fall sharply while output volume, productivity, and total value created rise – driven by compounding technological efficiencies rather than by demand contraction or scarcity. In the Inference Economy this signature emerges because inference-powered orchestration simultaneously cheapens the cost of work and expands the amount of economically viable work that can be performed.

To understand the core dynamic, examine the two distinct forces that operate together.

First, the cost side is mechanically deflationary. Compute is cheaper than human work for programmable tasks. The logic breaks when compute becomes more expensive than human work. What matters for the macroeconomic regime is the rate at which compute itself becomes cheaper. Inference costs are not a single number. They stratify by capability tier:

Commodity inference: The cost of achieving a fixed level of performance, such as what GPT-3.5 delivered at launch, has declined at a median rate of 200x per year for commodity-tier inference from January 2024 onwards^{277Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, March 12, 2025}, according to Epoch AI. It reflects hardware improvements, algorithmic advances (sparse architectures, mixture-of-experts), optimization techniques (quantization, distillation, dynamic batching), and scale economics.

Frontier inference: The cost of running the most capable model available at any given moment declines more slowly because each generation of frontier model is larger, denser, and more expensive to operate. The frontier advances. What was frontier inference 18 months ago is commodity inference today. But the cost of staying at the frontier does not decline at the same rate as the cost of a fixed capability tier. A NeurIPS 2025 Workshop paper using Epoch AI data (Cottier et al. and subsequent MIT work)^{278arXiv:2511.23455, The Price of Progress: Price Performance and the Future of AI, November 2025} found that the price to achieve a given level of frontier performance on knowledge, reasoning, math, and software-engineering benchmarks was declining 5-10× per year^{279Jean-Stanislas Denain, "How persistent is the inference cost burden?," Epoch AI, February 16, 2026280Stanford HAI, "AI Index Report 2025," Stanford University, 2025281Artificial Analysis, AI Model Benchmarks and Pricing}, with roughly 3× attributable to algorithmic efficiency after isolating hardware price declines^{282Hans Gundlach, Jayson Lynch, Matthias Mertens, Neil Thompson, "The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference," arXiv:2511.23455, November 28, 2025 (v2: March 23, 2026)}. Epoch AI’s February 2026 analysis independently confirms the 5-10× figure.

The blended cost trajectory for a diversified AGNT declines at a rate somewhere between the two curves because it runs a mix of commodity and frontier inference across a portfolio of workflow types. The precise rate depends on the orchestrator’s workflow mix and its ability to route tasks to the appropriate capability tier. This routing intelligence is itself a competitive advantage under the Third Law: the orchestrator that learns which tasks require frontier capability and which can be delegated to commodity models spends less for equivalent output quality. Anthropic demonstrated this dynamic in its April 2026 Advisor Strategy on the Claude Platform^{283Anthropic, The Advisor Strategy: Give Sonnet an intelligence boost with Opus, April 9, 2026; AWS, Introducing Anthropic’s Claude Opus 4.7 model in Amazon Bedrock, April 2026 (published April 17, 2026 per update tag)}, a deliberate inversion of the classical sub-agent pattern. A commodity-tier executor (Sonnet 4.6 or Haiku 4.5) drives the agentic loop end-to-end; Opus 4.6 is called on-demand as an advisor only when the executor encounters a reasoning wall. In Anthropic’s published benchmarks, Sonnet-with-Opus-advisor improved SWE-bench Multilingual from 72.1% to 74.8% while reducing cost per task by 11.9%; Haiku-with-Opus-advisor more than doubled BrowseComp accuracy (19.7% to 41.2%). The Advisor Strategy is a routing architecture that captures frontier quality at commodity-tier prices.

The deflation compounds regardless of tier. In year one, a workflow migrates from human execution to agent execution. In year two, the same workflow costs less because the relevant inference tier has cheapened. In year three, less still. The unit cost of orchestrated work declines structurally – exponentially for commodity workflows, meaningfully but more slowly for frontier-dependent workflows.

Three conditions could reverse the cost advantage and halt the ratio’s ascent. First, energy constraints: If power availability fails to scale with inference demand, the physical cost of producing tokens rises faster than algorithmic efficiency lowers it. Part I established that inference generates continuous baseload demand, that grid interconnection timelines run three to five years, and that the agentic transition cannot outrun the power grid. The Orchestration Ratio rises at a speed governed not only by inference economics but by the megawatts available to produce the tokens the economy demands.

Second, demand concentration: If agent populations grow faster than inference supply, pricing power shifts from the consumer of tokens to the producer, and the cost advantage narrows.

Third, regulatory intervention: If governments impose compute taxes, carbon costs, or usage restrictions that raise the effective price of inference above the cost of human labor in specific jurisdictions.

The second force driving deflationary growth is found on the revenue side. Deflation alone would mean shrinking revenue: the same work performed at lower prices. That is not what the Inference Economy produces. AGNT grows through two levers that operate independently or in combination.

The first lever is value. AGNT does not sell tokens. It sells outcomes: a claim adjudicated, a route optimized, a credit assessed, a shipment delivered on time. The outcome is priced against the value to the customer, not the inference consumed in producing it.

The Orchestration Ratio governs the cost of producing the outcome. It does not govern the price of the outcome. As the ratio rises and inference replaces human labor, the cost of production falls. If the outcome price holds, anchored by the context, trust, compliance, and coordination intelligence that do not deflate with compute, then the margin widens.

This is the structural mechanism of deflationary growth: costs decline with the Orchestration Ratio while revenues hold or expand through the value of the outcome. The orchestrator’s margin over raw inference cost is the orchestration “markup”. It rests not on the cost of inference, which is a commodity, but on the irreplaceability of the context and coordination intelligence that makes inference effective.

A company whose markup rests on thin context sees its margins erode under competition. That company fails the Second Law. AGNT’s margin is durable precisely because it is grounded in the same context that makes its orchestration position defensible.

The second lever is throughput. As inference costs fall, the volume of work that is economically viable to orchestrate expands. AGNT’s total addressable market grows with every reduction in inference cost. The prerequisite here is the existence of unbounded demand, where we go back to traditional economics. Augmenting the throughput only matters if you have a market for that.

§

What the Inference Economy Requires

The regime holds as long as three conditions hold. Each is currently met. None is guaranteed. The Orchestration Ratio is the metric that tells you which way it’s moving.

The Inference Economy is a regime, not a forecast. It endures only while inference remains cheaper than human cognition for programmable work, reliable enough for production, and embedded in systems that convert intelligence into priced outcomes at scale.

Each condition is currently met. None is guaranteed to persist indefinitely. The Orchestration Ratio is the metric that tracks the regime’s health. It rises when conditions hold, stabilizes when they weaken, and declines if they reverse. Tracking the ratio by sector, by workflow type, and by geography will tell the investor more about the pace and distribution of the agentic transition than any earnings call or product announcement.

The regime is real, but its pace remains uncertain. It is precisely in the gap between the structural endpoint and the transition path that both risk and opportunity reside.

Chapter 18 turns from the regime to its investment expression. The two value curves every company carries, the embedded option the market is not pricing, and why the instruments designed for the knowledge economy cannot detect the transition until it has already repriced the stock.

The market has begun to price what the agentic transition destroys. It has not begun to price what it creates.

§

The Embedded Option

Every company in the agentic economy carries two possible futures inside a single stock price. The gap between them is an embedded option. It runs in both directions.

On February 3, 2026, Thomson Reuters lost ~18% of its market capitalization in a single trading session^{284The Dupree Report, Anthropic Claude Triggers $285B Software Stock Crash, February 5, 2026}, the largest intraday decline in the company’s history. The catalyst was modest: an open-source legal plugin for Claude Cowork released four days earlier that automated contract review and NDA triage. A GitHub repository, built in weeks by a coding agent.

On the same day, ServiceNow, with more than 8,800 enterprise customers, $600 million in AI product ACV, 98% renewal rates^{285ServiceNow, 2025 Annual Report}, and the workflow control position that the Three Laws identify as the foundation of orchestration value, traded down alongside Thomson Reuters265, swept into the same sector-wide selloff.

It was unclear whether one company’s moat was eroding, whether the other’s Orchestration Layer was strengthening, whether both were vulnerable, or whether neither was. The market could see the immediate price reaction. It could not see the underlying dynamics with any precision. And it lacked the instruments to distinguish between those possibilities.

That failure is not a quirk of a single week. It is a structural property of the current moment. Understanding it requires a concept that has not yet formalized. Every company in the agentic economy carries two possible futures inside a single stock price.

Value Curve #1 is the fundamentals curve: the company as it exists today, growing at its current trajectory, operating within its current margin structure, valued by every DCF model and comparable screen in active use. It is precisely calibrated. It is what the market prices.

Value Curve #2 is the orchestration curve: the company after the crossing into the Inference Economy, delivering outcomes priced against the labor budget and beyond, compounding coordination intelligence with every workflow processed. No standard instrument contains a variable for it.

The gap between these two curves is an embedded option. It runs in both directions.

The positive option exists in every company with a plausible path from its current position to orchestrator status. For a deep-context incumbent with credible orchestration execution, the payoff may represent a doubling or tripling of enterprise value. This option is systematically unpriced because the instruments were designed for Curve #1 and contain no field for Curve #2.

The negative option exists in every company whose structural moat is already eroding beneath an income statement that has not yet registered the damage. Revenue is growing. Retention is stable. Margins are healthy. The DCF looks sound.

But the architectural position as measured by the Three Laws has already deteriorated. The company appears to be delivering Curve #1 at full value. It is not. The moat has broken. The revenue will follow. And the terminal value will evaporate. The instruments will detect the break only when it arrives as a financial result, by which time the negative option has already been exercised against the shareholders.

These are real options in the financial sense, not metaphors. Each has a probability of exercise, a payoff distance, a strike price determined by the investment required to cross, a time to expiry set by the competitive window, and a volatility determined by the uncertainty of the transition’s pace.

The central investment problem of the Agentic Era is to size both options, to see where the market prices upside at zero, and where it ignores downside because the income statement still looks clean.

§

The Negative Option: Moats Breaking in Silence

The more dangerous mispricing is the one that looks like stability while the terminal value is at stake.

Consider a vertical SaaS company holding fifteen years of customer data, charging $150 per seat per month. In Q1, revenue grows 12%. Gross retention is 95%. The DCF confirms a $40 billion valuation. Every instrument says the business is sound. None of them are asking whether the moat is intact.

In Q2, an orchestrating platform begins routing agents through the company’s API by reading its data and executing its workflows without any human opening the application. The data is still essential. The system of record still holds the truth. But it has been demoted from the system of action to the system of storage, from the place where work is directed to the place where data is queried. The customer’s employees stop logging in. The relationship between the company and the workflow it once controlled is severed, invisibly, at the architectural level.

In Q3, the customer asks why it pays $150 per seat for a product its employees no longer open. In Q4, the renewal becomes a repricing event. By Q1 of the following year, the revenue confirms what the architecture revealed twelve months earlier.

The moat broke in Q2. The revenue broke in Q4. The model detected it in Q1 of the following year. For twelve months, the company carried a negative embedded option. Its structural value was below the value the income statement displayed and the instruments saw nothing.

It is the mechanism that the February 2026 selloff exposed violently but imprecisely. Clearly, some companies were the proof that negative options are real. But the selloff’s indiscriminate nature reveals how poorly the market identifies which negative options are real.

The negative option is real for workflow wrappers and thin AI applications. These are companies whose moat was a subscription interface around functions a foundation model can now replicate. It is real for cognitive intermediaries: brokers, advisors, and consultants whose value resided in the programmable work that agents perform at a fraction of the cost. It is real for any company whose position on the Three Laws has deteriorated: intent captured elsewhere, agents routing around the system of record, coordination stalled.

It is not real for companies whose context is deep, whose intent position is secure, and whose orchestration execution is accelerating.

§

The Positive Option: Value the Market Cannot See

The positive option is the mirror image and it takes longer to price because its evidence is harder to detect.

When a company builds the Orchestration Layer, accumulates coordination intelligence, captures the intent surface, and begins delivering outcomes through the Inference Economy’s margin structure, the resulting value creation is gradual, unfamiliar, and largely invisible to instruments calibrated for Curve #1.

Consider Salesforce. Between late 2024 and early 2026, the company executed three moves whose cumulative logic is legible only if you are looking for orchestration. It repositioned Slack as an intent surface by rebuilding Slackbot as an agentic MCP client^{286Yahoo Finance, Salesforce lags in 2026 as software stocks react to AI concerns, January 15, 2026} that routes work to downstream agents, operates across the desktop, and converts conversational threads into structured CRM records without anyone opening Salesforce.

It acquired Informatica^{287Salesforce, Salesforce Completes Acquisition of Informatica, November 18, 2025}, absorbing the enterprise world’s leading data catalog and metadata management platform. This provided the context infrastructure that allows agents to reason across fragmented systems. And it scaled Agentforce to $800 million in ARR, growing 169% year-over-year, with 2.4 billion agentic work units delivered to date across Agentforce and Slack (“AWUs”)^{288Salesforce, Salesforce Delivers Record Fourth Quarter Fiscal 2026 Results, February 25, 2026},^{289Salesforce Form 8-K FY26 Q4, SEC filing}, a concept it specifically introduced for the Agentic economy, delivered and every one of its ten largest Q4 deals including agentic products.

The architecture is visible: Slack captures intent. CRM data and Informatica’s metadata fabric supply context. Agentforce coordinates the work between them.

Is this a company defending a CRM franchise? Or is it a company constructing an orchestration position across the enterprise revenue stack, and beyond, from its CRM base?

The skepticism has foundation. Organic revenue growth implies 7-8% for fiscal 2027, on a $41.5 billion base^{290Salesforce, Salesforce Delivers Record Fourth Quarter Fiscal 2026 Results, February 25, 2026; Salesforce Form 8-K, Q4 FY26 Earnings Release; The Motley Fool, Is Salesforce Stock a Buy After a Strong Earnings Report?, February 26, 2026}. The question is whether deceleration reflects maturity or transitional cannibalization from a faster-growing agentic business not yet at offsetting scale. There is genuine uncertainty about whether Agentforce will command the orchestration graph or merely participate in it.

But these risks are already priced. None of the above had prevented a roughly 30% year-to-date decline in Salesforce’s stock (as of April 10th, 2026)^{291CoinCentral, Salesforce (CRM) Stock Down 30% in 2026, April 2026}, because the market applied the same sector-wide negative-option framework to Salesforce that it applied to companies whose moats are genuinely breaking.

The consensus saw a CRM vendor whose per-seat model would compress under agentic substitution. The positive option, the probability-weighted value of Curve #2, in which Salesforce transitions from the system where records sit to the surface where work begins, was absent from the market price.

Orchestration revenue was growing inside the company. The AWU was a liminary proxy to apprehend the growth, yet to be translated into valuation assumptions.

This is the nature of the positive option. It hides in plain sight because it arrives dressed in the language of the old paradigm: revenue growth, margin improvement, customer expansion. The system records it has a modest continuation of Curve #1 rather than the early signature of Curve #2.

The accumulated expertise of the analyst becomes the obstacle. The more precisely calibrated the model, the more fluently it absorbs new evidence into old categories, the more powerfully it prevents recognition of what has changed.

The instruments are not broken. They are measuring the wrong regime.

§

The Three Mispricings

The two-sided option framework produces three categories of mispricing.

From this framework emerge three recurring mispricings in the Agentic Era, each defined not by what the market sees, but by what it fails to recognize.

The unrecognized long. A company carrying a large positive option that the market prices at zero. Its Curve #1 valuation may itself be under pressure from the sector-wide selloff, but the architectural assessment shows a company whose intent position is strengthening, whose context is deepening, whose orchestration execution is accelerating. The market sees the sector. It does not see the company. The positive option is free.

The unrecognized short. A company carrying a large negative option that the income statement has not yet revealed. Revenue is growing. Retention looks stable. The DCF produces a clean valuation. But the moat is breaking beneath the financials. Intent is captured elsewhere, agents are routing around the system of record, the seat-compression cascade is approaching. The market is paying full price for a Curve #1 that is already impaired. The negative option is unpriced.

The phantom orchestrator. A company the market has priced for Curve #2 is paying an orchestration premium that lacks the architectural position to achieve it. The earlier analysis of Palantir in this manifesto examined the clearest case in detail: extraordinary context depth, no intent proximity, and a valuation implying orchestration that the Three Laws suggest the architecture cannot deliver. The embedded option is negative, but the stock is priced as though it were positive.

These three mispricings generate longs, shorts, and pair trades in equal measure. They can be identified in advance not through financial data, which lags, but through structural assessment, which leads.

§

Losers First, Winners Later

There is a temporal asymmetry that governs how these options resolve, and it is the most important structural feature of the present moment.

The negative options resolve first. When an agent routes around a system of record, when a plugin replicates a workflow, when a comparison tool commoditizes an intermediary’s cognitive arbitrage, then the bypass is immediate, the revenue impact follows within quarters, and the market reprices in sessions. The old instruments can measure it because the categories already exist.

The positive options resolve later. Building the Orchestration Layer requires accumulating context across thousands of workflows, proving reliability, earning institutional trust, and navigating the seat-compression cascade as pricing shifts from seats to outcomes. Each crossing produces new orchestrators, but the financial signatures emerge gradually, quarter by quarter, as the Orchestration Ratio rises and Curve #2 separates visibly from Curve #1.

This temporal asymmetry explains the shape of the market in early 2026. The dispersion that reached historic levels was the negative options resolving, fast, violent, and sector wide. Software fell more than 20%^{292StockAnalysis.com, IGV Index commentary, February 2026}. Insurance brokers, wealth managers, real estate services, and freight brokers were hit in sequence across a single week.

The positive options had barely begun to resolve. The early financial signatures of Curve #2 are visible only within a few companies, and they remain small relative to the businesses they will eventually transform. The market has not yet learned to read them as leading indicators of a structural shift. They could not see the signs of the early value they may also reveal.

The dispersion will subside gradually. Not because the transition slows, but because the market learns to sort. The primitive, clumsy between-sector trade of the first wave gives way to the precise within-sector trade of the second. The destruction was priced in days. The creation will be priced in years.

Chapter 18 maps where the positive options are largest, including which domains cross next, at what pace, and what emerges on the other side.

The Coding Wedge conceals an orchestration Harness built for every domain. The programmable frontier it opens expands continuously. And the industries that cross last, the Leviathans, carry the deepest context moats, the widest embedded options, and the largest transformation surfaces in the economy. The crossing sequence inverts what the market expects about where value accrues.

§

The Consensus Story and What Lies Beyond It

Generative AI’s first economic impact is visible, measurable, and largely priced in. The market has rewarded the companies that use it to compress costs. The agentic crossing is a different phenomenon. This second economic impact is where the structural value sits.

C.H. Robinson is the clearest public-market proof point. The largest freight broker in North America reduced its workforce by ~31% from its 2022 peak to the end of 2025 (~11% YoY in 2025) while growing truckload volume 3% in Q4 2025^{293C.H. Robinson, SEC Form 8-K Q3 2025 Earnings Release, October 29, 2025}. Operating income rose 18.8% on revenue that fell 8.4%^{294C.H. Robinson / FreshPlaza, C.H. Robinson posts US$16.2 billion revenue in 2025, February 4, 2026}. Adjusted operating margin in the North American Surface Transportation segment reached 36.4%, up from 33.3%^{295FreightWaves, C.H. Robinson’s AI strategy boosts 2026 financial forecast, October 30, 2025}, with management targeting 40%298. The company automated 250 to 500 hours of manual tasks daily (December 2025 Investor Day)^{296Trucking Dive, CH Robinson CEO discusses technology, M&A, freight cycle, January 3, 2025} and achieved double-digit productivity gains in both its domestic and global forwarding segments298. CEO Dave Bozeman described NAST as being in the ‘third inning’ and Global Forwarding in the ‘first inning’ of its AI transformation and projected continued double-digit productivity increases through 2026. The strategy has a name: “Lean AI”, combining automation, artificial intelligence, and process redesign to decouple headcount from volume.

Walmart tells the same story on a different scale. More than 60% of Walmart’s U.S. stores now receive freight from automated distribution centers (targeting 65% by the end of FY2026)^{297Supply Chain Dive, Walmart grows automation usage throughout supply chain, November 20, 2025}, reducing unit delivery costs by approximately 20%^{298Supply Chain Dive, Walmart touts savings from continued automation efforts, May 8, 2025}. AI-powered systems in distribution centers predict and route disruptions, manage inventory placement, and erase repetitive manual tasks. EPS rose 26% in FY 2025^{299Walmart, Q4 FY26 Earnings Release, February 19, 2026; Macrotrends, Walmart EPS 2012-2026}. The company moved from the NYSE to Nasdaq, a deliberate rebranding to signal a tech-first identity. Walmart’s global advertising business grew 46% in FY2026 to $6.4 billion, with Walmart Connect U.S. up 41% in Q4^{300Walmart, Q2 FY2026 earnings presentation; Motley Fool, Walmart (WMT) Q4 2026 Earnings Call Transcript, February 19, 2026}. The margin expansion is driven by scaling high-margin services on top of AI-powered operations.

These are real results. The market rewards them. Analysts call for more. Every earnings call across the S&P 500 now includes a question about AI cost savings, and every management team has an answer prepared. The consensus is established: generative AI reduces costs, improves productivity, and expands margins.

This is Curve #1. It makes the existing business more efficient. It does not change the business. The company that stops at cost reduction earns a better margin on the same revenue model. The company that builds the Orchestration Layer changes the revenue model itself. This chapter asks how the Orchestration Layer extends from the domain that proved it – coding – into the enterprise at largethe domain that proved it – coding – into the rest of the enterprise.

§

The Wedge and Its Widening

Part III established why software engineering was the entry point. The question now is different: is the wedge still a wedge, or has it already begun to open?

Code provided the three conditions agent autonomy demanded: structured decomposition, objective verification, and high economic leverage. That made coding the natural entry point for every lab seeking to move from model provider to orchestration platform. What has happened since is that the labs have already begun extending the same Harness into adjacent domains, under product names that hide the common architecture beneath.

As discussed earlier, in January 2026, Anthropic launched Claude Cowork, an agentic assistant for non-developers^{301Fortune, Anthropic launches Cowork, a file-managing AI agent that could threaten dozens of startups, January 13, 2026}, running the same underlying architecture as Claude Code with different tools loaded into the Harness. Weeks later, it shipped domain-specific plugins for financial analysis, private equity, wealth management, and HR operations. Just a few weeks later, in April 2026, Claude Design launched with visual prototyping, slide decks, wireframes, and one-pagers generated conversationally. Figma’s stock fell as much as ~7% on the same trading session [Figure 18]. Claude Design connects directly to Claude Code, creating a seamless workflow from concept to shipped product without leaving the orchestration Harness.

The pattern is visible. Each extension loads different tools into the same Harness. Each triggers the same market reaction: an incumbent’s stock drops as the orchestrator extends its surface. The wedge is already widening. The question is how far and how fast it extends across the full economy and on what structural basis.

The answer begins with an architectural revelation that arrived by accident.

§

The Orchestration Graph

In March 2026, a build accident exposed the complete source code of Claude Code^{302The Register, Anthropic accidentally exposes Claude Code source code, March 31, 2026}. What the 512,000 lines reveal is not a coding assistant, but a domain-agnostic orchestration Harness whose underlying architecture is far more general than the product it currently serves. It is equipped with self-rewriting memory, an unshipped autonomous daemon, and multi-agent coordination capabilities still behind feature flags. The architecture does not merely describe how the crossing happens. It is the engine that drives it^{303Claude Code v2.1.88, exposed via npm sourcemap, March 2026. 785KB entry point, 40+ tools, multi-agent orchestration, consolidation engine ("Dream"), proactive daemon (KAIROS), and remote planning (ULTRAPLAN)}.

The tool registry contains over forty capabilities: file operations, web browsing, web search, task scheduling, cron job management, workflow script execution, push notifications, agent-to-agent messaging. A hook system with over twenty-five event types can intercept, deny, rewrite, or transform any action before or after execution. A remote planning runtime extends to thirty-minute horizons.

These are the primitives of general-purpose work orchestration: decompose, delegate, enforce, coordinate, verify, consolidate. None of these are specific to software engineering. The feature flags sealed into the build confirm the intent: WORKFLOW_SCRIPTS, COORDINATOR_MODE, DAEMON, VOICE_MODE. These are enterprise capabilities already engineered, gated behind compile-time switches, absent from external builds.

The convergence across labs confirms this is no accident. OpenAI’s Codex CLI uses a structurally parallel memory pipeline. Google’s Gemini CLI has added a memory manager agent^{304DeepWiki, google-gemini/gemini-cli page, last indexed April 2026}. Every major lab has converged on the same pattern: persistent memory with consolidation, multi-agent delegation, and tool registries that extend far beyond code. The shared decision: build a general-purpose Harness, deploy it where verification is tightest, then extend.

Three elements of this architecture have direct, structural implications for the crossing thesis.

The first is the multi-agent coordination system. It predicts which enterprise domains the Harness enters first. The sub-agent system supports three delegation models graduated by autonomy: bounded subtasks, persistent autonomous agents, and fully independent parallel workstreams. A Coordinator Mode spawns multiple workers for parallel research, centralized planning, distributed execution, and verification loops.

This is not a general-purpose design waiting for an application. It is a specific coordination topology – centralized delegation, parallel execution, and recombination – with a clear structural bias. It performs best on tasks that break into parallel components an agent can handle independently and then reassemble.

Financial analysis decomposes this way: revenue trends, cost structures, market comparisons, and regulatory checks. Each one is a bounded subtask, verifiable independently, and synthesized at the center. Insurance claims decompose this way. Compliance reviews decompose this way. The coordination architecture already built into the Harness is the one that the enterprise domains crossing first are structured to receive. This is not coincidence. It is why the transformation surface is largest where it is.

The second is the memory system. It is the mechanism by which verification costs fall. Every agent that runs long enough encounters context entropy^{305Luberisse, Verification Cost Asymmetry in Cognitive Warfare: A Complexity-Theoretic Framework, July 2025}: sessions accumulate noise, the model begins to contradict itself, and coherence breaks down. This is why the vast majority of production agents are capped at ten steps or fewer^{306arXiv, Measuring Agents in Production (Pan et al.), December 2, 2025 (v3 February 3, 2026)}.

The leaked architecture attacks this with a three-tier design: a lightweight pointer index loaded into every session (roughly 150 characters per entry, pointing to where knowledge lives rather than holding it), compaction strategies that continuously prune working memory, and a consolidation engine that runs during idle periods. A dedicated sub-agent gathers signals from recent sessions, merges new understanding with what is already known, resolves contradictions, and rebuilds the index to reflect current understanding rather than historical accumulation – a pile of historical records^{307Scortier, Claude Code Source Leak: What 512,000 Lines Revealed, April 2026; Milvus, Claude Code Memory System Explained: 4 Layers, 5 Limits, and a Fix, April 2026}. The system does not append. It rewrites. After twelve months of processing insurance claims, the consolidated memory is not a log. It is an evaluator with a structured model of how claims should be assessed, which patterns indicate fraud, which exceptions require human judgment, built from the operational history of doing the work.

The enterprise’s past operations, processed through the memory system, become machine-checkable ground truth. The Harness does not just perform work. It builds the verification infrastructure for the next unit of work as a by-product of performing the current one. This is why verification costs fall faster than the skeptic expects: the orchestration graph builds its own evaluator from the enterprise’s operational reality.

The third is the causal chain that connects orchestration to autonomy. It explains why domain expertise migrates to whoever operates the graph. Orchestration produces operational intelligence. Operational intelligence absorbs the domain expertise that was the barrier to autonomous operation. And the absorbed expertise is precisely what qualifies the agent to perform the work without human direction. One architecture. Two consequences.

The pointer memory that compounds understanding of a codebase compounds understanding of a claims portfolio, a loan book, or a logistics network with identical mechanics. What we call understanding gravity: the accumulated, cross-system operational intelligence that no individual application can capture. This belongs to whoever occupies the center of the graph.

And it is domain-specific. Understanding gravity accumulated in insurance does not transfer to logistics. Each domain’s orchestration graph must be built by the entity that possesses that domain’s operational reality.

This is why the Leviathans, the asset-heavy operators with decades of irreplaceable context, carry the widest embedded option in the crossing.

An orchestration graph is the architecture that emerges when this system becomes the center of an enterprise’s work. Through open protocols like MCP, the model reaches outward to CRM, ERP, data warehouse, and vertical tools. The systems of record become nodes. The model is the center. Each system sees only its own transactions. The orchestrator sees the workflow that connects them all. The coordination architecture determines which domains it enters. The memory system builds the verification infrastructure as it operates. The understanding gravity compounds in ways that reward early movers and punish those that wait. The engine is domain-agnostic. What differs across domains is the verification threshold – the cost of confirming the agent’s output is correct. And the skeptic’s challenge to the thesis rests entirely on whether that cost is falling.

§

The Verification Threshold

Code is deterministic. Enterprise work is not. The skeptic argues that the orchestration graph cannot extend beyond software because the rest of the economy lacks the verification signal that made code tractable.

Most enterprise workflows are not binary. They are probabilistic-verifiable: outcomes checkable against structured rules, historical patterns, institutional policies, and domain-specific criteria. An insurance claim either meets coverage criteria or it does not. A loan application either satisfies underwriting policy or it does not. A compliance filing either passes regulatory thresholds or it does not. A freight route either meets the delivery window at the contracted cost, or it does not.

These outcomes are verifiable. They are verified against domain logic, not a compiler. And that domain logic is increasingly machine-encoded. The question that determines the crossing is whether the cost of that verification is falling fast enough for agents to clear the same threshold code cleared.

Three forces are converging to collapse it.

The first is the orchestration graph itself building verification infrastructure as a by-product of its own operation. As established above, the memory system consolidates operational history into structured evaluators. But the effect extends beyond the Harness. Every claim adjudicated, every loan underwritten, every compliance check performed across the enterprise is training data for evaluation systems that encode institutional judgment in evaluable form. More than half of production agent teams already use LLM-as-judge. One model checks the output of another as first-pass verification, complemented by human exception review^{308LangChain, State of Agent Engineering, 2026}. The enterprise’s past operations are becoming the compiler for its current outputs. Tand he orchestration graph accelerates this by consolidating those operations into structured, queryable intelligence that the next verification cycle draws on.

The second is enterprise platforms restructuring as agentic ground. Salesforce’s Headless 360, unveiled April 2026, exposes every capability in its platform as an API, MCP tool, or CLI command so agents operate the system without a browser^{309Salesforce, Introducing Salesforce Headless 360. No Browser Required., April 15, 2026}. When the largest CRM in the world rebuilds to serve agents as primary users, verification infrastructure is embedded in the programmatic interface itself: policy enforcement, audit trails, compliance checks.

Last, formal verification in the LLM itself is extending beyond mathematics. Companies like Axiom train AI systems to produce outputs in the proof language Lean, where every reasoning step is machine-checkable and logically guaranteed correct^{310Menlo Ventures, AI Will Write All the Code. Mathematics Will Prove It Works., March 12, 2026}. Today, this applies to mathematical reasoning and code. The trajectory points toward quantitative finance, regulatory compliance, and pharmaceutical safety – any domain where the cost of an undetected error justifies formal proof. As verification becomes fast, cheap, and automatic, the boundary between deterministic and probabilistic collapses.

The threshold is falling. The Harness is ready. The question is where and how to size the opportunity.

§

The Transformation Surface

For the companies that cross, the prize is sized by how far the programmable frontier extends into their operations. The frontier is not static. It expands with every advance in context, verification, and workflow design. Understanding its structure requires distinguishing what is programmable today from what becomes programmable tomorrow.

Under Orchestration Economics, cost of goods sold for an orchestrated workflow is inference tokens, not human compensation, benefits, and management overhead. And inference costs deflate 5-10x per year^{311arXiv, The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference, November 28, 2025} (for frontier capability). The margin structure improves automatically as the substrate cheapens.

The Carnegie Mellon-Stanford research^{312arXiv, Wang et al., How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations, October 26, 2025 (v2 November 6, 2025)} provides the first rigorous framework for sizing the surface. It identifies three tiers of programmability that determine where the frontier sits today and where it is moving^{313Carnegie Mellon–Stanford programmability framework (2025). Three tiers: readily programmable (20–50% of tasks, 88.3% speed gain, 90.4–96.2% cost reduction), partially programmable (amenable but requiring workflow redesign or human judgment at decision points), minimally programmable (aesthetic judgment, relationship management, strategic oversight). Programming capability required in 82.5% of knowledge work occupations (U.S. DOL ONET data)}.

Readily programmable work includes deterministic tasks solvable through code execution: data analysis, system configuration, computational tasks, and structured writing. We estimate this tier represents 20-50% of total tasks across knowledge work occupations^{314arXiv, Wang et al., How Do AI Agents Do Human Work?, October 26, 2025}. For these tasks, agents achieve 88.3% speed gains at 90-96% cost reduction322,^{315PPC Land, AI agents complete human work 88% faster but with significant quality gaps, November 10, 2025}. The workflows are redesigned entirely around programmatic approaches.

Partially programmable work includes tasks theoretically amenable to programmatic approaches but requiring workflow redesign or human judgment at key decision points: complex engineering analysis, strategic planning, content creation, and operational optimization. This tier is where the frontier expands. Each advance in context length converts a partially programmable task into a readily programmable one. The analyst who spent four hours on a quarterly comparison was performing partially programmable work until the agent could hold the full quarter’s filings in working memory simultaneously. The associate team reviewing a hundred contracts was performing partially programmable work until million-token context windows made the entire corpus traversable in a single session. The orchestration graph’s memory system further accelerates this: as it consolidates domain-specific patterns, it builds the institutional intelligence that converts judgment calls at decision points into verifiable pattern recognition. The partially programmable tier is not a fixed residual. It is the expanding frontier.

Minimally programmable work includes tasks requiring aesthetic judgment, relationship management, negotiation, strategic oversight under irreducible ambiguity, and contextual understanding that current agents cannot replicate. This tier stays human. This is why the Inference Economy is not about replacing humans, but about reorganizing work per its programmable nature and re-organizing, wholly, the company as such. The seasoned underwriter assessing a novel risk. The site superintendent making a judgment call about ground conditions. The dealmaker reading a negotiation. Agents handle the readily and partially programmable tiers, freeing the domain expert to concentrate on the minimally programmable work where judgment creates the most value.

The three tiers map onto the crossing sequence that we develop after. Wave One industries have the highest proportion of readily programmable work: insurance claims processing, banking compliance, procurement, and customer support. Wave Two industries have more partially programmable work that is becoming addressable as context and verification advance: legal analysis, healthcare administration, complex analytics. Wave Three industries have the most complex mix – but they also carry massive readily programmable functions inside their cost structures (back-office, compliance, reporting) alongside the operations-bound, minimally programmable judgment work that defines their context moats.

The programmable frontier is therefore not a single threshold to be crossed. It is a surface that expands continuously, driven by three measurable forces: context length (how much domain information the agent can hold simultaneously), verification cost (how cheaply the enterprise can confirm agent output), and workflow redesign (how effectively the organization restructures around human-agent teaming). Each advance converts partially programmable tasks into readily programmable ones, widening the surface.

Evidence from agents in production confirms where the surface is deepest today. Finance and banking account for 39.1% of deployed production systems. This is the largest concentration of any domain^{316arXiv, Pan et al., Measuring Agents in Production, December 2, 2025; Lovelytics, Validating the Shift: How Lovelytics & Databricks Solve the Agent Reliability Paradox, December 16, 2025}. Technology follows at 24.6%, corporate services at 23.2%, data analytics at 13.0%, and legal, customer support, and healthcare services each at 8.7%^{312arXiv, Wang et al., How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations, October 26, 2025 (v2 November 6, 2025)}. The Google DeepMind/MIT scaling study provides the structural explanation for why finance leads^{317arXiv, Towards a Science of Scaling Agent Systems, December 2025; Google Research, Towards a science of scaling agent systems: When and why agent systems work, January 28, 2026; MIT Media Lab, Project Overview, 2026} and it maps directly onto the coordination architecture described above.

When tested across four agentic benchmarks – financial reasoning, web navigation, game planning, and workflow execution – centralized multi-agent coordination improved performance on the Finance Agent benchmark by 81%^{318Research.Google, Towards a science of scaling agent systems, January 28, 2026; MIT Media Lab, Project Overview, 2026}, the largest gain of any domain tested. Workbench (business task automation) showed marginal effects^{319ImagineX Digital, Why Your Multi-Agent AI System Is Probably Making Things Worse, February 18, 2026; DEV / evoailabs, Stop Blindly Scaling Agents: A Reality Check from Google & MIT, December 12, 2025}. PlanCraft (sequential planning) degraded by 39-70%^{320Google DeepMind/MIT, "Towards a Science of Scaling Agent Systems" (December 2025). 180 configurations across four benchmarks, three LLM families, five architectures. Finance Agent: centralized MAS +80.9% vs. single-agent baseline (mean 0.631 vs. 0.349). Domain complexity score: Finance 0.407 ("moderate decomposability; structured domains amenable to localised agent reasoning"). Framework predicts optimal coordination strategy for 87% of held-out configurations}.

The mechanism is the one for which the Harness was built. Financial tasks decompose into parallelizable subtasks: revenue analysis, cost structure review, market comparison, and regulatory checks. This allows agents to work independently and synthesize at a central coordinator. This is not a generic finding about AI capability. It is a finding about a specific coordination topology, centralized delegation with parallel execution and recombination, which is precisely the Coordinator Mode already engineered into the leaked architecture. The paper’s domain complexity framework identifies “moderate decomposability” as the structural sweet spot for multi-agent scaling^{321Emergent Mind, Scaling Agent Systems: A Quantitative Study, December 9, 2025}. Finance sits squarely in it. So do insurance, compliance, and corporate back-office operations, domains whose workflows break into verifiable components that agents process in parallel and recombine. The domains crossing first are the domains whose task structure matches the coordination architecture the Harness already possesses.

§

The Crossing Sequence

Three structural properties govern which industries cross in which order: output verifiability, the routine-to-judgment ratio, and latency tolerance. The three programmability tiers determine how deep the surface extends in each. The industries that score highest on all three properties and carry the highest share of readily programmable work cross first.

Wave One – Crossing Now (2025-2026). Insurance: claims adjudication against codified coverage rules, massive historical loss data, high routine-to-judgment ratio. Banking and financial services: credit analysis, KYC/AML, trade settlement, risk reporting – the domain where multi-agent coordination produces the largest measured performance gains. Corporate services: HR, procurement, invoice matching, expense management – the industry-agnostic cognitive overhead of any organization. Customer support: autonomous case resolution already demonstrated at 50% within twelve days of deployment^{322CX Today, Salesforce Turns Slackbot Into an AI Agent, Cuts Case Resolution, April 2026}.

Wave Two – Entering Production (2026-2028). Legal and compliance: contract analysis, regulatory monitoring, due diligence – structured documents that million-token context windows can hold simultaneously. Healthcare administration: prior authorization, billing, care coordination – the administrative layer consuming 25-30% of US healthcare spending^{323Center for American Progress, Excess Administrative Costs Burden the U.S. Health Care System, November 2, 2021; Trilliant Health, Hospital Administrative Expenditures Exceed Direct Patient Care by Nearly 2x, October 17, 2025; Fierce Healthcare, Wasteful administrative spending costs US healthcare up to $570B a year, October 10, 2022}. Data analytics: report generation, anomaly detection, ad hoc analysis.

Wave Three – The Leviathans (2027-2030). Supply chain and logistics: route optimization, demand forecasting, carrier negotiation, exception handling. Construction and engineering: scheduling, compliance documentation, change order management. Manufacturing operations: quality documentation, production scheduling, multi-site regulatory compliance.

Wave One assignments are directly observed in production deployment data. Wave Two has partial evidence combined with strong structural logic. Wave Three is structural extrapolation – these domains appear in the long tail of deployments, and the three properties predict their crossing as verification costs fall.

The sequence conceals the most consequential inversion in this chapter.

§

The Inversion

The market reads the crossing sequence as a timeline. Wave One is investable now. Wave Three is too far away to price. This reading is precisely backwards.

Crossing speed and option depth are inversely correlated.

Wave One industries cross first because their verification is most tractable. A claims processing workflow is rule-bound and learnable. An AI-native services company can replicate it without decades of operational history.

Wave Three industries cross later because their verification is operations-bound – checked against physical outcomes, environmental variables, and real-time conditions no rulebook fully captures.

This is the Leviathan thesis operating inside the crossing sequence. The carriers, insurers, banks, manufacturers, and pharmaceutical companies of Chapter 15 are the same companies that appear in Wave Three. Their operational context satisfies all Three Laws: proximity to intent, irreplaceable context depth, and workflow intelligence that compounds with every shipment routed, policy underwritten, and site managed. And the domain-specificity of understanding gravity – established above – means the orchestration graph for logistics can only be seeded by an entity that has routed a billion shipments, the graph for insurance only by an entity that has underwritten decades of real risk. The memory system compounds what it is given. What the Leviathan gives it is irreplaceable.

And here is what is likely ignored: Wave Three companies do not wait for Wave Three. They carry massive volumes of Wave One and Wave Two work inside their own cost structures. Every Leviathan has a back office. Every Leviathan has legal, compliance, analytics, and customer support. The Leviathan captures the productivity gains of Waves One and Two inside its own operations while building the Orchestration Layer on irreplaceable operational context.

The cost transformation flows from the waves already crossing. The defensibility flows from the context that takes longest to deploy but is hardest to replicate. The orchestration graph’s understanding gravity compounds at the center – and for the Leviathan, that center sits atop decades of operational reality no competitor can reconstruct.

The Leviathan that captures the readily programmable tier across its back-office functions (Waves One and Two) while building an Orchestration Layer on context no competitor can replicate (Wave Three) is not cutting costs. It is capturing Curve #1 savings on the readily programmable tier, expanding the frontier as partially programmable work becomes addressable, and simultaneously pulling back margin that had leaked to intermediaries and deploying its operational intelligence as commercial agents inside customer operations – moving from an industrial multiple of 8x toward platform economics. The minimally programmable remainder- the irreducible judgment that defines the Leviathan’s expertise – becomes more valuable, not less, because the Orchestration Layer has freed the domain expert to concentrate where human judgment creates the most value. The market sees proximity. Wave One is near, Wave Three is far. This manifesto sees defensibility. The industries that cross last carry the deepest context moats, the widest gap between Curve #1 and Curve #2, and the largest embedded orchestration option in the economy. The largest value creation of the Agentic Era will not occur in the companies building AI. It will occur in the companies that use it.

§

The anatomy of the crossing is complete. C.H. Robinson and Walmart demonstrate what cost reduction looks like when it works and why it remains Curve #1. The coding wedge has already begun widening through Claude Cowork, Claude for Finance, and Claude Design, each loading different tools into the same orchestration Harness. The leaked source code confirms the Harness was always domain-agnostic and its three structural elements drive the thesis forward: the coordination architecture predicts which domains cross first, the memory system manufactures verification infrastructure as a by-product of operation, and the understanding gravity it accumulates is domain-specific, compounding only for the entity that possesses the operational reality. The verification threshold is collapsing from three directions. The transformation surface is structured by three tiers of programmability – readily, partially, and minimally programmable – and the frontier between them expands with every advance in context, verification, and workflow design. The crossing sequence follows verifiability, decomposability, and latency tolerance, and is inversely correlated with defensibility.

The Leviathans carry the widest option.

What happens to value when the crossing is complete?

The Re-Rating

Everything so far changes the income statement. What follows changes the multiple. This may be the most consequential mechanism of value creation in the Agentic Era.

For a generation, the market assigned multiples based on financial architecture. An industrial at 8x because it was capital-intensive, cyclical, and slow-growing. An insurer at 10x. A bank at 12x. Software at 30x or 40x because it exhibited recurring revenue, high margins, low marginal cost, and operating leverage that expanded with scale.

When a company crosses into the Agentic Era, it will acquire structural properties that command a different multiple. Recurring outcome-based revenue replaces cyclical or transactional income. Structurally expanding margins replace thin and variable ones. Compounding coordination intelligence replaces static positioning.

And capital productivity, the return on trucks, policies, loans, equipment, improves without additional capital deployed. A fleet coordinated by agents achieves higher utilization without more trucks. An insurance book underwritten by agents prices risk more precisely without additional reserves. A loan portfolio managed by agents reduces defaults without reducing yield.

The industrial at 8x that begins to exhibit these properties re-rates to 15x or 20x as the market correctly recognizes that the financial architecture of the business has changed. This will seem jarring or speculative to those who don’t understand Orchestration Economics. But it should not.

A company generating recurring orchestration revenue, compounding coordination intelligence, and producing higher returns on the same capital base is a platform, regardless of whether it also owns trucks, policies, or loans.

The re-rating is larger for non-technology companies than for technology companies, because the distance between starting multiple and destination multiple is wider. An enterprise software company that improves from 8x to 12x forward revenue captures a meaningful but bounded re-rating. The market already understands software economics.

A global insurer that moves from 10x to 20x earnings has doubled its multiple. The market has never seen an insurer that looks like a platform. The surprise will be larger. The option is wider. The mispricing is deeper.

The compounding of higher earnings at a higher multiple is the mechanism. It will operate most powerfully in the 431 non-technology companies of the S&P 500^{324S&P Dow Jones Indices, S&P 500 Communication Services & Information Technology Index (93 constituents as of Jan 30, 2026)}. These are the companies whose physical assets, operational context, and customer relationships become more valuable, not less, when intelligence is abundant and coordination is the scarce resource.

What remains is to demonstrate our method for sorting.

The framework has now moved beyond theory and broad economics. It can now be applied across the index to reveal a clear distribution.

§

From Physics to Engineering

The three laws describe the forces. What follows is the instrument that measures them.

The Architectural Resilience Assessment Framework (ARAF™) was developed to take the next logical step beyond the Three Laws: convert structural analysis into a score, and a score into an investment signal.

§

The Architecture

ARAF™ measures two forces in opposition.

The first is Opportunity: four dimensions that determine where a company sits in the agentic economy.

O1: Intent Proximity. Where does the user’s goal originate? The company that sits at the point of intent formation controls routing. Everything downstream is interchangeable. O1 measures not whether a company touches intent, but whether it originates it. The difference between a search engine and a database is O1.

O2: Context Depth. What does the company know that cannot be replicated from public data? Ten million processed claims. A billion routed shipments. A trillion dollars of underwritten credit. O2 distinguishes between data – which can be licensed – and context – which must be earned through operational repetition.

O3: Workflow Control. Does the company direct the work, or merely participate in it? The distinction between a system of action and a system of record is O3. One tells agents what to do. The other stores what was done. The first is an orchestrator. The second is a filing cabinet.

O4: Technical Readiness. Can agents actually interact with this company’s systems? API coverage, protocol compatibility, authentication architecture, and real-time data access. O4 is necessary. It is not sufficient. This is where most conventional AI analysis stops. ARAF starts here and looks upward.

The first three dimensions define Layer Position: where a company sits in the stack. Orchestration Layer, context layer, tool layer, or infrastructure layer. Layer Position is the structural diagnosis. O4 is the technical readiness to act on that position.

The gap between the two is the Layer Gap. This is the single most important diagnostic signal in the framework.

A positive gap means a company has strategic assets trapped behind technical limitations. It is a Context Prisoner.

A negative gap means a company has technical excellence without strategic positioning. It is an Excellent Tool.

§

The second force is Risk – five vectors that measure how directly the agentic transition threatens the company’s current position.

R1: Bypass. Can agents route around the company’s interface entirely? If the user interface disappears, does the value proposition survive?

R2: Displacement. Can agents absorb the cognitive work the company’s users currently perform? The scissors attack: agents commoditize from below while asset owners absorb from above.

R3: Model Consolidation. Can a foundation model replicate the company’s functionality as a built-in feature? Every company one API call from obsolescence carries R3 exposure.

R4: AI-Native Competition. Are new entrants building the same capability without legacy constraints?

R5: Technical Debt. Does the company’s architecture resist the agentic transition? Monolithic systems, batch processing, and legacy authentication. Every architectural constraint that slows agent integration is R5.

§

The third element is AIR – the Agentic Implementation Reality check. It measures whether a company is executing on its opportunity, or merely presenting slides about it. Real agent deployments. Credible timelines. Production reliability metrics. AIR separates the companies that are building from the companies that are announcing.

§

ARAF produces a single output: the Adjusted Option Value. This is the probability-weighted capacity of a company to achieve durable orchestration, adjusted for the risks that could prevent it and the evidence of its execution. It is the embedded option that sits inside every stock in the index, waiting to be sized.

The framework sorts companies into five structural positions:

Context Prisoner

Excellent Tool

Platform Defender

Emerging Orchestrator

Vulnerable

Each has a distinct transformation path and a distinct risk profile. We examine two here.

§

The Context Prisoner

This is a species of company with extraordinary operational context trapped in legacy architecture.

Think of a large bank or financial institution. Consider that such global operators that have processed billions of transactions over decades. Each interaction compounds into proprietary operational intelligence: patterns, edge cases, and real-world behavior under stress. This context is immense and irreproducible. Models can’t learn it from data, and startups can’t recreate it without decades at scale.

This is where we find the potential to layer on new revenue: operational intelligence deployed as agents inside customer systems.

But the context sits deep inside systems built twenty or thirty years ago. Monolithic architectures. Fragile integrations. Data locked in formats that agents cannot consume. The AI engineering team, where it exists, is small relative to the legacy maintenance workforce. Implementation reality is low, not because leadership is unaware, but because the technical debt is measured in decades and the organizational transformation required is the most complex program the company has ever undertaken.

Meanwhile, the competitive window is closing. Model providers are building orchestration platforms that could query the company’s data and direct its systems without owning the context, capturing the orchestration surplus while the asset owner remains a rich but passive substrate. AI-native entrants are building lighter, faster versions of the same workflows, but without the tangible (or intangible) assets but without the technical debt either.

The Context Prisoner’s context is a wasting asset if the architecture that unlocks it is not built before the alternatives mature. The companies that solve this will re-rate from industrial to platform multiples, anchored in non-replicable assets. Those that do not risk seeing the value of their own operations captured by a competitor or by an orchestrating layer beyond their control.

§

The Emerging Orchestrator

The second species is a company building the Orchestration Layer in real time, with measurable velocity, but whose moat is young.

The archetype is a technology company, though not exclusively, that has secured a strong workflow control position and is accumulating context at speed. It captures intent across thousands of enterprise customers. Its agentic products are shipping, adopted, generating retention deltas that confirm the value is real. R&D allocation is high and rising. Implementation reality is among the strongest in its peer group. The revenue layer is already forming: outcome-based pricing, agent-to-agent interactions, customers restructuring their operations around the platform’s orchestration capabilities.

What makes this profile distinct, and what makes the investment case both compelling and fragile, is the relationship between speed and depth.

The Emerging Orchestrator is winning the race for orchestration position in its domain. But the context it has accumulated is measured in months and years, not decades. The behavioral patterns, the exception paths, the relational intelligence that compound into an irreplaceable moat are forming, but they have not yet reached the density that makes them structurally defensible against a well-resourced incumbent that decides to move. The risk is specific and real.

A Context Prisoner that wakes up and unlocks its decades of operational intelligence and deploys it through an Orchestration Layer arrives in the same domain with context the Emerging Orchestrator cannot match. The incumbent’s data is deeper, its customer relationships longer, its domain expertise more granular. If the incumbent moves before the Emerging Orchestrator’s moat thickens, the younger company’s early lead may prove to be just that: early, not durable.

The history of technology transitions is populated with first movers who built the category and second movers who captured it.

The Emerging Orchestrator’s defense is speed and the Third Law. Every workflow orchestrated deepens its coordination intelligence. Every customer onboarded adds context. Every quarter of compounding widens the gap between what the platform knows and what a new entrant could learn.

The question is whether the compounding rate exceeds the rate at which incumbents mobilize. In some domains, the Emerging Orchestrator has already accumulated enough context to be defensible. In others, the race is genuinely open and the outcome depends not on capability but on the institutional velocity of legacy organizations that have never moved at this speed.

The market tends to price Emerging Orchestrators on revenue growth, a Curve #1 metric that captures the trajectory but not the structural position. The embedded option is

in the durability of the position, which depends on how quickly context compounds relative to the competitive window.

A company growing at 25% with a context moat thickening faster than incumbents can respond carries a larger positive option than a company growing at 40% whose orchestration position is replicable. The growth rate is visible. The moat trajectory is not. Unless you have the instruments to measure it.

§

The Excellent Tool Trap

These are companies that score high on technical readiness but low on the strategic dimensions, such as intent proximity, workflow control, and hub position that determine who orchestrates and who is orchestrated.

Layer Position is the weighted average of the first three and tells us where the firm sits in the agentic stack. Technical Readiness is separate and reveals whether the firm can execute there. The Layer Gap is the Layer Position minus O4. This is the diagnostic.

A negative gap wider than 0.20 means the firm occupies a lower ring than its engineering would suggest. It is technically strong and structurally mispositioned. This is the Excellent Tool Trap.

The trap is that this feels like progress. The product ships, usage rises, retention improves, and sales celebrate. The metrics move. The category does not. A calendar that drafts invitations is still a calendar. A CRM that suggests next steps is still a CRM. A code editor that writes functions is still a code editor. The tools become excellent. They do not become orchestrators.

The Palantir case reveals that an orchestration position is not a single attribute. It is a conjunction. A company must sit close to where intent originates. It must accumulate context that compounds with use and cannot be replicated. And its coordination learning must accelerate in ways that create asymmetric switching costs. Architecture enables all three. Architecture guarantees none.

The Excellent Tool Trap is a category error that infects every technology valuation built on the assumption that architectural quality equals strategic positioning. If the market can misprice a company this aggressively by ascribing orchestration premiums to a context substrate with excellent tools, then it is almost certainly making the same error elsewhere.

The trap applies equally in both directions. A SaaS incumbent that bolts AI features onto a legacy product is in it. So is an LLM lab that ships a Harness-wrapped copilot and calls it orchestration. The test does not care who built the product. It cares what position the product occupies in the workflow. A coding agent that a developer still invokes per-task, per-prompt, remains a tool regardless of how capable the underlying model is. The labs that extend from Ring 2 into Ring 3 are not exempt from the Excellent Tool Trap. They face a version of it shaped by their own adjacency to intelligence: the temptation to mistake a more capable Harness for a captured workflow.

§

The Agentic Sorting

The agentic transition is not a technology sector event. Of the five hundred companies in the S&P 500, seventy-nine sit entirely outside it. Gold miners. Regulated electric utilities. Self-storage landlords. Their value is determined by commodity prices, rate-setting commissions, and the physical scarcity of land. Agents do not change the geology, the regulation, or the rent.

Another thirty are AI infrastructure: semiconductors, data center REITs, and optical networking. Their fortunes depend on demand for intelligence, not on their own capacity to orchestrate it. They are the substrate, not the species.

Strip those away and 391 remain. That’s 78% of the index. The Three Laws sort these companies into six positions.

We have classified every one of them against the Three Laws by measuring Layer Position (the combination of intent proximity, context depth, and workflow control) against Technical Readiness (the capacity for agents to actually interact with the company’s systems). The distribution that follows is the result.

§

Context Prisoners: 167

These are the companies with the deepest operational context in the economy – and the widest gap between what they know and what an agent can access.

Banks with four decades of credit underwriting data running on batch-processing mainframes. Insurers with ten million claims in systems designed before the browser. Pharmaceutical companies whose clinical trial libraries, adverse-event databases, and regulatory correspondence exist nowhere else. Defense contractors whose maintenance records for fighter jet engines span thirty years and three wars. Consumer goods companies whose route-to-market optimization was built one truck stop at a time across a billion deliveries.

One hundred and sixty-seven companies. Nearly a third of the index. Their Layer Position is high – they sit close to intent, hold irreplaceable context, and in many cases still direct the work. But their Technical Readiness is low: agents cannot consume their data, interact with their systems, or orchestrate their workflows without a modernization program that will take years and cost billions. The layer gap is the defining feature.

That gap is the largest source of mispriced optionality in the market. The companies that close it will re-rate from industrial multiples to platform multiples, anchored in non-replicable assets. Those that do not risk seeing the value of their own operations captured by an orchestrating layer they do not control.

§

Platform Defenders: 91

These already possess both moats – and no gap between them.

Payment networks processing billions of authorizations per day through real-time, API-native rails. Enterprise software platforms whose workflow data spans a hundred and fifty thousand corporate customers and two decades of process history. Financial data companies whose indices, ratings, and risk models are wired into the plumbing of global capital allocation. Retailers whose fulfilment systems coordinate warehouse robotics, last-mile delivery, and demand forecasting inside a single stack. Leading insurers whose telematics data and modern underwriting platforms put them on the right side of the gap. Money-center banks that spend more on technology annually than most software companies earn in revenue.

Layer Position is high. Technical Readiness is adequate or better. Context and technology, present in the same company. The question is not whether they can transition from system of record to system of action. It is whether they will – fast enough – before the API-era leverage they currently enjoy becomes a wasting asset.

The market prices many of them as mature. The Three Laws say they occupy the most contested ground in the economy.

§

Ascending Tools: 31

Technically superb companies – security platforms, developer infrastructure, design automation – that have recognized the Excellent Tool Trap and are building their way upward. Their Technical Readiness is high. Their Layer Position is rising: acquiring context, constructing intent surfaces, expanding from executing functions to directing workflows.

These are not orchestrators today. They are companies whose O1, O2, and O3 are demonstrably increasing – through acquisitions that add context, through product expansions that capture intent, through platform moves that shift them from function execution to workflow coordination. The trajectory is the bet.

§

Tool Layer: 46

The mirror image. Technically excellent companies whose Layer Position is low and not rising. They score high on Technical Readiness – API-first, agent-compatible, usage-based pricing – and low on the strategic dimensions that determine who orchestrates and who is orchestrated. They provide infrastructure or functions that orchestrators call. They do not capture intent. They do not accumulate irreplaceable context. They do not direct workflows.

The market frequently confuses technical excellence with strategic positioning. A company can score 0.85 on Technical Readiness and still be a commodity tool. The Excellent Tool Trap is not about capability. It is about layer.

These companies are structurally exposed. When the Orchestration Layer matures, they will compete on price for a function call. The market has not priced this.

§

Emerging Orchestrators: 19

This is the rarest position in the index because the S&P 500 selects for incumbents, and orchestrators are by definition challengers. Their Layer Position is high and their Technical Readiness is high – but they are building the position, not defending one they inherited.

The deeper pipeline is private. Anthropic, OpenAI, SpaceX, and a generation of private-market companies waiting in the IPO queue. When that pipeline surfaces, the valuation gap between orchestration-layer economics and infrastructure-layer economics will become impossible for the market to ignore.

§

Vulnerable: 37

Workflow wrappers with no proprietary data. Cognitive intermediaries caught in the scissors – agents commoditizing expertise from below, asset owners absorbing function from above. Thin applications one model-provider plugin away from irrelevance.

Low Layer Position. Low Technical Readiness. No clear transformation path.

§

The six positions produce a structural ratio.

Not every AGNT candidate becomes an orchestrator. Context Prisoners could fail to modernize – the technical debt is measured in decades and the competitive window is closing. Platform Defenders could protect the system of record too long and find the orchestration positions occupied. Ascending Tools might never reach the Orchestration Layer. Emerging Orchestrators might see their moats thickened by incumbents that move faster than expected.

The structural math, applied with transformation probabilities across the six categories, produces a distribution:

Approximately 125 emerge as durable orchestrators – Context Prisoners that unlock their assets, Platform Defenders that complete the transition, Ascending Tools that reach the Orchestration Layer, and Emerging Orchestrators whose moats prove durable.

Approximately 266 face value compression – Context Prisoners that fail to modernize, Platform Defenders that defend too long, the entire Tool Layer competing on price for a function call, and the Vulnerable dissolving.

Twice as many losers as winners. The dispersion between them is the largest structural opportunity in the index.

This manifesto provides the vocabulary. The scoring framework provides the measurement. The instruments to exploit the dispersion are the work that follows.

Machines became actors.

It is a discontinuity: a structural break in the relationship between human work and machine capability that invalidates the categories, valuation frameworks, and institutional assumptions built for the knowledge economy.

The crossing occurred when machines moved from processing instructions to pursuing goals by receiving intent, decomposing tasks, selecting tools, executing autonomously, and delivering outcomes.

The evidence is measured and reproducible: in software engineering, where AI now authors a material share of production code across the world’s most sophisticated technology companies; in professional services, where frontier models perform equity research, legal analysis, and financial modeling at the level of trained professionals; in enterprise operations, where agents are deployed in production across domains, beyond the pilot phase and producing actual revenue and actual margin gains.

The paradigm shift did not require universal adoption. It occurred when the capability became proven, reproducible, and commercially available. That moment has passed. The technology required to exploit the shift already exists. Intelligence is sufficient. Agents are deployed. Infrastructure is being built at historic scale. The transition is no longer theoretical. It is operational.

When machines become actors, value no longer accrues to those who supply intelligence. It shifts to those who orchestrate it. This is the central thesis of this manifesto.

The Orchestration Layer becomes the new locus of value. Not the model alone. Not the interface alone. Not the data alone. The position that receives intent, applies context, coordinates execution, and compounds learning across workflows becomes the strategic control point of the new economy. Three Laws determine who holds that position durably.

Proximity to intent determines who owns the human-work relationship. Context depth determines whose position is defensible. Coordination compounding determines whose advantage strengthens with scale. These laws do not merely describe software. They describe the reorganization of value across software, services, and the physical economy.

The sorting they imply is already in motion. In software, systems of action separate from systems of storage. Tools with semantic density survive as throughput nodes while wrappers and thin intermediaries erode. In services, cognitive middlemen are compressed as agents absorb the comparison, synthesis, and routing work that once justified entire fee pools. In the real economy, physical asset owners and operational leviathans begin ascending into the Orchestration Layer above the infrastructure they already control.

This is the deeper inversion. For a generation, markets rewarded asset-light business models and treated physical infrastructure as drag. In an economy of abundant cognition, the scarcest resource becomes everything that is not cognition: real assets, regulatory positions, operational memory, institutional trust, and the context generated only by actually running the world.

That is why the largest value creation of the Agentic Era may not occur inside the technology sector. It may occur in the enterprises that own the physical networks, balance sheets, claims histories, supply chains, and installed bases on which orchestration can compound.

What the crossing produces is not simply faster software or cheaper labor. It produces a new economic order.

Revenue shifts from access to outcomes.

Costs shift from labor to inference.

Moats shift from contractual friction to operational intelligence.

Capital productivity rises as the same assets generate more output under intelligent coordination.

The result is a new financial species: companies whose earning power improves across revenue, margins, and capital productivity simultaneously, because the same underlying force reshapes all three at once.

This is why the old frameworks fail.

They were built for a regime in which intelligence was scarce, labor was the throughput constraint, and capital productivity moved slowly. They assume stable relationships between headcount and output, between assets and returns, between software margins and market size. Those assumptions no longer hold. The discontinuity is not that one variable changes. It is that many variables that used to move independently now begin moving together.

That does not mean the transition is riskless, orderly, or immediate. It is not.

The pace remains uncertain. Model convergence may slow. Reliability thresholds may take longer to cross in high-stakes domains. Enterprises may resist outcome-based pricing. Regulation may delay deployment. As Kuhn understood, the greatest friction in any paradigm shift is often not technological but institutional: old frameworks persist because the people and organizations built around them persist.

There is also a more immediate danger. Being right about the direction of the paradigm does not protect anyone from being wrong about its pace. This transition is being financed through one of the largest infrastructure buildouts in modern economic history. If demand arrives more slowly than the capital stack requires, the destruction of capital may be severe even if the structural thesis is correct. The history of technological revolutions is full of investors who were right about the future and wrong about the timing.

That distinction matters. But it does not reverse the central argument.

The direction is clear. Machines are increasingly becoming actors. Orchestration is how value reorganizes around that fact. The companies, sectors, and institutions closest to intent, richest in context, and able to compound coordination across real workflows will increasingly capture the value that abundant intelligence releases.

The framework offered here is not a prophecy. It is a lens. It will be tested, refined, and in places corrected by evidence. But the sorting it describes is not a metaphor. It is arithmetic applied to a discontinuity.

The knowledge economy was organized around the storage, retrieval, and distribution of information.

The next economy will be organized around the orchestration of work. The institutions that understand that crossing first will not simply outperform within the old order.

They will help define the new one.

To my team at Decoding Discontinuity – Eric, Thibaud, Fouad, Nicolas, and Nicolas – for the endless work put into this Manifesto. Thank you for bringing invaluable new perspectives to this work.

To Rob and Michael, who backed research that did not fit any conventional brief and trusted me early in my journey. That kind of confidence is rare.

To Marco and Gordon, for the fascinating conversations and for challenging the views in this Manifesto. Our discussions were key in refining the thesis.

To Raj, for allowing me to see what agentic transformation looks like from inside a company that moves the physical world. My vision of the Leviathans would not have been fully formed without those exchanges.

To Marc and John, for their early enthusiasm for the ARAF framework and for their trust in my vision.

To Véronique, who translated a structural argument into images with infinite patience. Her hands and taste are everywhere in these pages.

To Chris, who supported me in this journey from the start and prevented me from making this text even longer. Thank you for pushing me at all times and for forcing me to clarify my thinking, even when I resisted.

To Mathieu, Arthur, and Stanislas. They never asked me to stop, instead encouraging me to continue at all times. I owe them a lot.