The Company Brain Era: When Skills Become Employees, Agents Become Labor

At the 2026 Y Combinator Startup School, Garry Tan told a story about a 17th-century philosopher who ground lenses, then turned the topic to himself: he said his code output this year was 400× what it was in 2013. The reason wasn't that models got smarter, but that he decomposed his work into four pieces—Skill Files that spell out "how to do it," a Resolver Table that decides "who does it," a Company Brain that remembers "what's been done," plus a "rented" model intelligence that drops in price every quarter and can be swapped out at any time. These four terms, put together, are really answering a bigger question: when labor can be hundreds of concurrently running agents, does the organization itself still need to look the way it does today?

I. One Talk, Four Metaphors

Where this talk happened. At the 2026 Y Combinator Startup School, Garry Tan delivered a keynote titled "Own Your Intelligence." The official description reads: "YC's president and CEO on working memory, Skill Files, and why your rented intelligence can never compound like intelligence you own." This 42-minute talk was later fully transcribed (full transcript here), dissected into ten clips and widely circulated; it is the primary first-hand source for all analysis in this article. A few months earlier, at the more engineering-oriented AI Engineer World's Fair, Tan gave a companion talk—on how to rebuild the company structure itself "the Agent way." The framework from that talk was organized into a comparison table by Trustabl.ai, which we will use in Section II.

Tan opened with a winding historical analogy: the 17th-century philosopher Baruch Spinoza was excommunicated from his Jewish community at 23, and thereafter made a living grinding lenses by day and writing dangerous books by night. Tan drew on Spinoza's heterodox claim that "God is Nature, diffused through all things" to analogize today's misunderstanding of AGI—everyone is waiting for a singular "announcement of arrival" moment, but he believes true general intelligence has already appeared in an unassuming form: "It's not like a god, it's like infrastructure: a terminal window, a markdown folder, a job that finishes running while you sleep." He calls this Personal AGI—not "general intelligence shared by everyone," but "general intelligence belonging to one person."

The real skeleton of the talk is four interlocking metaphors, scaling from the individual to the organization:

1. Models are rented (Rented Intelligence). "A frontier model is rented, a commodity, and it gets cheaper every quarter; plus your own, unique context; plus a harness that connects the two—that's the formula for the next decade. The quality of the model is rented, but your brain is ideally your own."
2. Skill Files are employees. A markdown document that spells out "exactly how to do this one thing," detailed enough that "a smart intern could follow it," is an executable, encapsulated "employee" of specific capability.
3. Resolvers are org charts. A task-dispatch table that decides which Skill File a new task should go to—essentially the digital version of "who reports to whom, who's responsible for what" in a traditional organization.
4. Company Brain is institutional memory. A knowledge base shared by all agents, continuously written to and retrieved from, so that every task execution "knows everything the company already knows" without re-explaining the context each time.

None of these four terms is new in isolation—RAG, knowledge bases, task routing, prompt engineering are all familiar to AI practitioners. What's worth examining deeply is that Tan positions them on the coordinate system of "organization": he explicitly says, "When you sit down tonight to work with an agent, you're not writing code, you're managing a workforce made of markdown." That sentence is the origin of this article's title.

II. Redrawing the Org Chart: Skill File, Resolver, Workforce Made of Markdown

Unpack Tan's metaphors and you find he's essentially renaming a set of existing technical components in the familiar language of startups:

This comparison table comes from Trustabl.ai's distillation of Tan's AI Engineer World's Fair talk, and is highly consistent with the Startup School talk—showing this isn't a one-off analogy but a framework Tan has repeatedly refined and crystallized over the past half-year.

What's interesting is Tan's criterion for a "Skill File": not whether it technically runs, but "if a smart intern could follow this instruction, then an agent can execute it." This criterion democratizes "writing down how to do something" from an engineer's monopoly to something anyone can do—in the talk he gave an example: a YC finance colleague with no engineering background folded ~100 Excel worksheets into an agent-driven internal tool. "She's not a programmer; she's now an agent manager." This is also the most easily verifiable and most viral part of the narrative: markdown is the new code, and the compiler is the language model itself.

But behind the redrawn org chart lies a harder operational distinction Tan calls "latent space vs. deterministic space": taste, judgment, "understanding what a vague request really wants"—this kind of computation should stay in the model, guided by a markdown file; while arithmetic, scheduling, and database queries should be handed to deterministic code. His example is scheduling a 6,000-person conference—"having an agent or a person arrange seats for five people is easy, latent space can handle it; but building personalized schedules for 6,000 people, the latent-space agent must call code." This distinction isn't rhetoric; it's what he identifies as the root of "almost every agent failure case"—giving arithmetic that belongs in code to the model's "intuition," or conversely, using rigid code to handle fuzzy requests that require judgment.

III. Company Brain: From Personal Memory to Delegable Organizational Memory

Tan's personal knowledge base is called G-Brain, which he describes in the talk as a "Karpathy-style" markdown knowledge wiki containing about 220K pages—"25 years of life, diarized." Every email, every meeting, every draft is indexed, retrieved, and cross-referenced. The system has been open-sourced, and one of its key tutorials is directly named company-brain.md—an operations guide for scaling from "personal brain" to "company brain."

This document lays out the leap from personal to organizational very concretely; the core is three structural changes:

• Multiple sources: Meeting notes are one source, each colleague's own client notes are another, the company shared wiki is a third—they live in the same database but remain independent.
• Per-user OAuth scoping: Each team member has their own credentials, which determine which sources they can read and write—"Alice writes her own client source, reads her own plus shared; Bob writes internal ops source, reads his own plus shared; neither can see the other's writes." The docs emphasize the team has done "fuzz testing across every read path, zero leaks."
• Personal folders and dedicated Skills: The shared brain has a shared structure, but each person also has their own subfolders, their own scheduled tasks (weekly reports, client follow-ups), and their own restricted Skills.

This design answers a challenge Tan himself raises in the talk: a brain with no maintenance "is just a landfill with search—retrieval will present outdated information with full confidence; a badly written Skill File will permanently固化 a bad process." His solution is to "treat the brain as production infrastructure": every piece of information is annotated with its provenance, conflicts between new and old information trigger verification, and there's a dedicated "pruning" role—in some sense, the agent-ified version of the traditional "knowledge management" position, except instead of passive document archiving, it's an active, living memory continuously read and written by agents.

It's worth calling out one layer here: Company Brain matters not because it's a better search engine, but because it changes the cost curve for "new employees" (whether human or agent) entering the organization. In a traditional org, a new person takes weeks to months to "know what the company knows"; in Tan's architecture, any newly spun-up agent, once connected to Company Brain, "knows everything the company knows" from second one. This is the key to defining Company Brain as "institutional memory" rather than mere "knowledge base"—it fulfills the function that previously required org charts, apprenticeship systems, and handoff documents.

IV. The Numbers Speak: Revenue Per Employee Breaks Historical Scales

If the previous two sections remained at the level of architectural metaphors, the numbers in this section are the most verifiable—and most challengeable—part of this narrative.

Tan gave two YC-affiliated company examples in his talk: Emergent (S24 batch) went "from public launch to nine-figure revenue in just 8 months," crossing the $15M ARR mark with 15 employees, roughly $1M per person; Retell AI (W24 batch) hit $60M ARR with ~40 people, a similar ratio but at 3× the scale. Tan's exact words: "This revenue per capita has never existed before—not in software, not in oil, not in railroads."

Update (2026-09): The Emergent figures Tan cited are a snapshot from when it crossed the $15M ARR milestone in fall 2025. Per TechCrunch on July 15, 2026, Emergent has closed a $130M Series C, valuation rising to $1.5B within five months, achieving unicorn status; ARR rose to $120M (70% growth in four months), and headcount expanded from 15 to ~200 (mostly in Bangalore). Revenue per employee thus fell from ~$1M at the $15M ARR crossing to ~$600K—still about 2× the traditional SaaS benchmark, but no longer the extreme outlier. This neatly confirms the caution in Section VII: such per-capita revenue figures are more like snapshots of a company at an explosive moment than a steady state that can be linearly extrapolated.

These two figures aren't isolated cases. Zoom out and you see several corroborating data points:

• TechCrunch reported that the Swedish "vibe coding" platform Lovable broke through $400M ARR in February 2026 with just 146 full-time employees—adding $100M in revenue in a single month. Per capita, that's ~$2.7M, exceeding Google (~$1.6M) and Meta (~$1.9M) revenue per employee, while the traditional SaaS benchmark typically sits in the $200K–$400K range. Gartner had predicted "by 2030, a new generation of unicorns with $2M revenue per employee will emerge"—Lovable achieved it four years early.
• Forbes analysis shows AI-native startups generally fall in the $2M–$4M revenue-per-employee range, 6–10× the traditional SaaS mean; image generation tool Midjourney generates ~$200M revenue with ~11 employees, ~$18M per person—the most extreme case in this set of figures.

Update (2026-09): Lovable's curve hasn't flattened. Per TheNextWeb on June 9, 2026, Lovable published its first "build economy" data report: ARR climbed from $400M in February to $500M, headcount still at 146. Compared to other companies in this high-per-capita cohort (Emergent later expanded to 200 people), Lovable is so far one of the few samples that keeps pushing the "revenue per employee" curve up rather than diluting it back toward the traditional benchmark as it scales—but the report also cautions that these figures are self-reported, unaudited, and that annualized subscription revenue can retreat at any time.

On the enterprise side, Nvidia CEO Jensen Huang gave a more direct staffing ratio at GTC 2026: "Over the next decade, we want to keep headcount at about 75K... those 75K employees will work alongside 7.5M agents"—100 agents per person. Huang frames this as "growth" rather than "replacement"—but looked at another way, this is the world's most valuable company officially putting "agent count" into the numerator of its workforce planning.

How to read these numbers matters more than the numbers themselves. Three layers of caution are necessary: first, almost all of them come from companies' own public disclosures or founder interviews, with no independent audit; second, these are surviving extreme cases, not industry means—Tan himself acknowledges this in the talk: "I know what correlation is, so I want to be careful: I can't prove that AI-generated code caused the growth." Third, "revenue per employee" as a metric is historically distortion-prone in early-stage high-growth SaaS companies (one-off large customers, front-loaded customer-acquisition spend not yet reflected on the cost side), and whether AI-native companies will repeat this trap is too early to call. But even after all these discounts, the fact that in the YC W25 batch, "a quarter of the companies have codebases that are 95% AI-generated, and this batch is on track to be one of the fastest-growing and most profitable in YC history"—organizational data at this scale is hard to dismiss entirely with a one-liner about "survivorship bias."

V. An Old Question Reopened: Coase's Theory of the Firm

If this discussion stopped at "AI improved productivity again," it wouldn't be more consequential than any efficiency revolution of the past decade. What makes it worth taking seriously is that it touches an 89-year-old question in economics: why do firms exist?

In 1937, Ronald Coase answered in "The Nature of the Firm": because market transactions have costs (finding people, negotiating, contracting, monitoring). When "the cost of organizing an additional transaction within the firm" falls below "the cost of completing the same transaction through the market," the firm will internalize that transaction, expanding until the two costs equalize. This logic has dominated organizational theory for nearly a century—firm boundaries are, in essence, boundaries drawn by transaction costs.

In 2026, a crop of economics and computer science papers has independently revisited this question:

• Economist Gillian Hadfield (a long-time researcher in AI governance and law-and-economics) and Andrew Koh, in their paper "An Economy of AI Agents," pose the AI version directly: "If a significant fraction of transactions begin to be carried out by AI systems, what happens to firm boundaries? How does the large-scale introduction of AI agents change organizational and industrial structure?" One of the paper's conjectures: AI agents may drastically reduce coordination costs, enabling a single firm to simultaneously hold capabilities previously dispersed across multiple markets, triggering a phase transition from "many specialized firms" to "a few giant firms spanning multiple industries."
• Tassilo Klein and Sebastian Wieczorek's paper "The Headless Firm: How AI Reshapes Enterprise Boundaries" offers a more concrete model: in traditional modular systems, integration cost grows quadratically with the interaction topology (O(n²)); in a "protocol-mediated" agent system, integration cost collapses to linear growth (O(n)), while verification cost scales with task throughput rather than interaction count. This shift selects for an organizational form they call the "Headless Firm"—shaped like an hourglass: a highly personalized generative interaction interface at the top, a layer of standardized protocols in the middle, and a competitive market of highly specialized execution agents at the bottom.
• The more radical paper "Structural Dissolution" (Chao Li and Chunyi Zhao) directly challenges Coase's premise: Coase's framework assumes "coordination must happen between distinct economic agents," with firms and markets as two options for "who can do this cheaper"; but when coordination can be completed inside a computational system rather than between agents, this "Coasean choice set" itself collapses—the firm may continue to exist as a legal entity, but the "coordination" function it originally performed is being ceded to regionally governable AI infrastructure. A similar argument appears in Michael Harre and Paul Ormerod's "The Coasean Reversal," except with the opposite conclusion—they argue that the complexity of frontier technology production has exceeded the coordination capacity of external market clusters, which will instead make firms more inclined to internalize capabilities previously obtained through market collaboration.

Read these papers together and you find the academic community currently has no consensus, but the axis of debate has shifted from "will AI improve efficiency" to "when coordination costs are drastically lowered by agents, how much reason for existence does the firm as an organizational form still have?" And Tan's Resolver Table—a routing table that decides "which Skill File a task should go to"—is precisely a concrete, runnable specimen of "how coordination gets automated" in this theoretical debate: when the org chart no longer needs HR department approval but is a piece of version-controllable routing logic, the "internal coordination cost" that Coase calculated is on a completely different ledger.

VI. Who Owns Judgment: The Property Politics of Skill Files

Midway through the talk, Tan abruptly shifted tone from "efficiency gains" to "a political question," telling a fictional but precise story.

A customer-support engineer named Maya spent two years teaching her agent 40 skills: how to handle a P0 incident at 2 AM, how to soothe a customer about to churn, how to write a postmortem that actually prevents the next incident. Those 40 files are the judgment she accumulated over two years.

Version 1: These files live in Maya's own codebase. When she changes jobs, the files go with her. Each additional year, her judgment compounds another year—this is "ownership."

Version 2: These files live in the company's codebase, governed by company IT policy. When Maya leaves, she takes nothing; "the company continues running a version of her that will never ask for a raise." Tan's exact words: "She doesn't have a career; she's experienced an extraction."

"I believe the Skill File is yours. Own your own skills—because if you don't own it, your job becomes a skill file."

Tan connected this back to industrial history: artisans once owned their own tools, which preserved their independence; factories broke that—the loom belonged to the factory, not the weaver. Knowledge workers thought they were safe because their tools (i.e., knowledge and experience) lived in their heads, where no one could confiscate them. "The Skill File, for the first time, makes your cognition extractable, archivable, versionable, ownable—the only question is, ownable by whom."

This passage is worth analyzing on its own because it transforms the proposition "agents become labor" from a purely technical proposition into a property-rights proposition—and property rights have historically been the sharpest entry point for labor politics. Placed in a larger coordinate system, this is actually a new version of an old problem: knowledge workers' "tacit knowledge" previously couldn't be fully encoded, so it naturally belonged to the individual, and companies could only indirectly "rent" that knowledge through the employment relationship; the emergence of Skill Files makes tacit knowledge extractable, copyable, and capable of running independently of its creator for the first time. This is both the fundamental source of efficiency gains in this system (judgment no longer drains away with employee turnover) and its greatest political risk—once judgment can be fully packaged and run detached from its creator, the boundary between "employment" and "licensing" becomes blurry: is the company employing Maya herself, or a perpetual license to Maya's judgment? Current employment contracts, non-compete agreements, and work-for-hire IP assignment rules are almost entirely unprepared for this question. It's foreseeable that "who owns the Skill File on departure" will likely become a new, specific negotiation clause in Silicon Valley employment contracts over the next few years, just like today's non-compete and equity-vesting clauses.

VII. Dissenting Voices: Survivorship Bias and the Collapse of the "Coworker" Narrative

Before accepting Tan's narrative wholesale, at least three layers of critique deserve a place on the table.

The first layer is the word "coworker" itself. A year before Tan's talk, two Forrester analysts were already debating whether the "AI agent as coworker" framing holds. Skeptic Anthony McPartlin listed five reasons, the sharpest being "false equivalence": establishing some kind of parity between agents and real employees isn't just misleading, it's diminishing to the practitioners involved—"'coworker' implies a collaborative relationship; human workers in Amazon warehouses or Tesla factories surrounded by robots never call those robots 'coworkers,' so why should sales roles?" He also notes that this narrative is naturally advantageous for companies selling AI products—"when a Salesforce-level company redefines its long-term value on very limited evidence, buyers should be wary."

The second layer is the dilution of the "one-person company" myth itself. Even outlets that present "revenue per employee" most compellingly acknowledge that the extreme cases currently circulating are more "proofs of concept" than replicable industry benchmarks—most cited claims of "one-person companies hitting a billion dollars" don't withstand scrutiny, closer to inspirational narrative than verifiable financial disclosure.

The third layer is one Tan himself acknowledges: he repeatedly emphasizes "I can't prove that AI-generated code caused the growth, this is just correlation"; he also proactively raises "a brain with no maintenance is a landfill with search"—the retrieval system will present outdated or even incorrect information with full confidence, and a badly written Skill File will permanently固化 a bad process. This means the efficiency gains of this system are highly dependent on continuous human curation—which is precisely the part that's hardest to scale, hardest to quantify, and easiest for startup narratives to ignore. In other words, if you push the "Skill File = employee" metaphor to its extreme, the company doesn't need less judgment, but needs judgment concentrated in fewer, more critical people—the organizational hierarchy gets flattened, but the responsibility and risk borne by the small handful of "brain curators" at the top of the pyramid is actually amplified.

VIII. A New Anatomy of the Organization: If the Company Becomes an Hourglass

Assemble the evidence from the previous sections and you can see a more specific organizational-form hypothesis than mere "efficiency improvement," one that neatly aligns with the "Headless Firm" hourglass model in Klein and Wieczorek's paper:

• Hourglass top: A highly personalized generative interaction interface—in Tan's language, the founder or core decision-maker who possesses the Company Brain and can "know everything the company knows." Tan's exact words: "You're the founder, and you're the entire management layer."
• Hourglass middle: A layer of standardized protocols—corresponding to the Resolver Table, determining how tasks flow between Skill Files; essentially converting the HR department's "who should own this" function into version-controllable routing logic.
• Hourglass bottom: A market of highly specialized execution units that can be competitively replaced—corresponding to individual Skill Files / agents, which can theoretically be swapped out at any time like outsourcing a specific function today, replacing an "underperforming employee."

The most radical implication of this shape isn't "companies get smaller," but that the middle layer gets compressed to near zero. In traditional organizations, middle managers' core value is precisely "coordination"—translating top-level strategy into specific tasks and aggregating execution-layer information back to the top. When the Resolver Table and Company Brain respectively take over task routing and information aggregation, middle management's previously irreplaceable "translation" function has, for the first time, a technical substitute. This is why Klein and Wieczorek's paper emphasizes "verification cost"—an hourglass organization doesn't eliminate supervision, but shifts the supervised object from "people" to "task throughput," requiring an entirely new, outcome-oriented rather than process-oriented performance system. Tan's "Trigger Evals" is the most rudimentary prototype of this system.

For teams considering whether to move in this direction, a few relatively actionable criteria:

1. Separate "latent space" from "deterministic space" before automating. Tan repeatedly emphasizes that "almost every agent failure case" stems from giving arithmetic that belongs in code to the model's judgment, or vice versa. Draw this boundary wrong, and all subsequent organizational redesigns will be built on quicksand.
2. Treat "company brain" as production infrastructure, not an appendage of a knowledge-base project. A knowledge base without continuous pruning, provenance tracking, and conflict detection propagates errors faster and more confidently as it scales.
3. Before expanding the team, write "who owns the Skill File on departure" into the contract. This isn't an edge clause that can wait until there's a problem—it determines whether the company is willing to let employees truly commit their judgment to the system, and it determines the company's future attractiveness in the talent market.
4. Use revenue-per-employee figures for directional sense, not budget assumptions. They're currently closer to existence proofs of extreme values than linearly extrapolable industry benchmarks.

IX. Conclusion: Whose Intelligence, Whose Memory

Return to the ungainly formula from the opening of this article: "A frontier model is rented, a commodity, and it gets cheaper every quarter; plus your own, unique context; plus a harness that connects the two." If this holds, then the true competitive dimension among future organizations will rapidly shift from "whose model is stronger" to "whose Company Brain has been accumulated longer, pruned cleaner, delegated more precisely"—because the gap at the model layer is being erased on a quarterly timescale, while the accumulation of organizational memory and judgment is a moat that competitors can almost never replicate overnight.

But the most unfinished part of this narrative is precisely its most political part: when a company's core assets shift from "which people we employ" to "which Skill Files and which version of Company Brain we own," the ancient compact between employee and organization—"I sell you my time, but my judgment and experience remain my own"—has, for the first time, the possibility of being technologically dismantled. Maya's story is the quietest yet most widely circulated passage in the entire talk because it doesn't stop at the exhausted question of "will AI take jobs," but asks a sharper and sooner-to-land question: when your work can be fully written into a file, who should that file belong to?

This question currently has no answer. But it's certain that in the coming years, the ones answering it won't be just engineers and founders, but also labor-law scholars, unions, and every ordinary employee who, for the first time, is asked to "write down how you do this as a Skill File."

Key References

Garry Tan, "Own Your Intelligence" talk, Y Combinator Startup School 2026 (full transcript) · Garry Tan, GBrain open-source project · company-brain.md · Garry Tan on Building AI-Native Companies, Trustabl.ai · Ruben Dominguez, Garry Tan Runs YC at 400x His 2013 Output, The AI Corner · Lovable adds $100M in a month with 146 employees, TechCrunch · Paul Baier, AI-Native Firms Lead In Revenue Per Employee, Forbes · Jensen Huang on NVIDIA's 100-agents-per-employee plan, Fortune · Indian AI coding startup Emergent becomes a unicorn with $130M Series C, TechCrunch (2026-07-15) · Lovable hit $500 million in revenue with 146 employees, TheNextWeb (2026-06-09) · Anthony McPartlin, The "AI Agent As Coworker" Narrative Is Nonsense, Forrester · Gillian K. Hadfield, Andrew Koh, An Economy of AI Agents, arXiv:2509.01063 · Tassilo Klein, Sebastian Wieczorek, The Headless Firm, arXiv:2602.21401 · Chao Li, Chunyi Zhao, Structural Dissolution, arXiv:2604.27435 · Michael Harre, Paul Ormerod, The Coasean Reversal, SSRN

Author: Feng Xiaoping, [email protected]