The self-improving company.

The short version.

In May 2026 Y Combinator published a thirteen-minute batch talk by Tom Blomfield, one of its General Partners and a co-founder of Monzo and GoCardless, called "How to Build a Self-Improving Company with AI".¹ It argues that most companies are still organised like a Roman legion, with people passing information up and down a hierarchy, and that AI makes that design obsolete. In its place he proposes a company built as loops that sense, decide, act, check and learn, running without a person in the path so that the company "starts to self-improve even when you're sleeping".

The talk draws on Jack Dorsey and Roelof Botha's essay "From Hierarchy to Intelligence", published a few weeks after Block cut its staff from over 10,000 to just under 6,000.⁵⁷ It has been widely shared, and it is sometimes described as an Anthropic lecture. It is not: Blomfield gave it as a YC partner, and announced in July 2026 that he was taking leave from YC to join Anthropic's compute team.⁹

I study it here because boards are now being shown this picture, and parts of it are right. The useful ideas are making a company's know-how legible, treating internal software as disposable and keeping people for the moments that need judgment. The weak parts are the ones a board would sign off: loops that merge and deploy their own code with no named person, token use as a measure of effort, recording every conversation, and removing the managers who develop people. On the published record, two of the talk's ten main claims hold, four partly hold, two are not shown and two run against the evidence.

Status as of 25 October 2026: YC has published no design, review rules or results for the internal agent described in the talk, and no data for its revenue-per-employee figure. Every time-sensitive claim on this page was checked on that date.

Where the talk came from.

The talk sits in a run of related work. On 26 February 2026 Block announced it would cut from over 10,000 people to just under 6,000, about 40% of staff.⁷⁸ Dorsey later described sizing the company at the fewest people who could keep the service running, meet regulatory commitments and meet Wall Street's, and said mistakes had been made in the calculation.⁵⁶ On 31 March he and Roelof Botha of Sequoia published "From Hierarchy to Intelligence". It traces corporate structure to the Roman army, "a nested hierarchy with a consistent span of control at every level", and concludes: "There is no need for a permanent middle management layer."⁵⁶

Within YC, Diana Hu published a Startup School talk on 24 April on building a company with AI from the ground up, and Garry Tan has open-sourced GBrain, a Markdown knowledge base that takes in meetings, email and messages.³¹⁷¹⁶ Blomfield credits both. He also mentions "Codex 5.5", which is OpenAI's GPT-5.5, released in Codex on 23 April.¹⁰ Those references place the recording in late April or early May 2026. YC published it on 21 May.¹²

What the talk argues.

Blomfield opens with the legion. Rome projected power from the centre to Hadrian's Wall through nested units with fixed spans of control, and named people who passed orders down and information back up.¹ (He places the wall in Scotland; it runs across northern England.¹⁹) "If you think about most companies today, they are organised like a Roman legion, where human beings are the conduit for information flowing up and down."

He then rejects last year's frame for AI: copilots that make engineers 20% more productive, assistants added to existing workflows. That is "taking the old way of working and adding a more powerful engine onto it". The alternative starts by extracting the company's domain knowledge, which lives in people's heads, Slack, email and Notion, and turning it into shared context and skills. Once that is legible to AI, he argues, a company can be rebuilt as "a set of recursive, self-improving AI loops".¹

How the loop works.

The talk's loop has five stages, which Blomfield credits to Diana Hu. A sensor layer takes in signals from the world: customer emails, support tickets, code changes, cancellations, product telemetry. A policy layer sets what the system may do, what needs a person's permission and what must be logged. A tool layer gives it deterministic APIs. A quality gate checks its work with evals, deterministic checks, safety filters and "human review for high-risk stuff". A learning mechanism finds where it failed and feeds that back to the top.¹³ "If you can run every single step of that without human intervention, your system gets better and better while you're sleeping."

Two details matter for what follows. The talk's own quality gate keeps human review for high-risk work, which sits awkwardly with running every step without a person. And Blomfield is candid about how far anyone has got: "I'm not sure anyone has a truly self-improving company in every function."¹

What YC says it built.

The example in the talk is an internal agent at YC. It began as a tool staff could ask, with deterministic access to YC's database: when did I last hold office hours with this company? It then became a sidekick, using retrieval and several query routes to suggest five founders in petrochemicals for an introduction. Blomfield calls that stage last year's AI, "making me 20 or 30% more effective".¹

The change came with a monitoring agent that reads every query every YC employee makes, sees which failed and asks why: is a new tool needed, a new skills file, a database view, an index? Then, in his words: "Let's write the code, put in a merge request to the YC codebase, have an agent review it, and merge it and deploy it." The next day the same query succeeds. "This literally happens overnight now."¹

He describes two more loops in the same pattern. One reads product analytics, finds the step in the sign-up funnel with most friction, runs an A/B test for a week, keeps the winner and repeats. The other triages customer suggestions with agents acting as chief product officer and chief technology officer, and ships the ones in line with the roadmap overnight, "without a human being involved".¹ None of these systems has been published. Everything in this section rests on the speaker's account.

The five implications.

Burn tokens, not headcount. Blomfield says YC companies now reach Demo Day "with about five x more revenue per employee than they did 18 months ago", and that companies will soon be "constrained on token usage, not on headcount". He would measure each person's token use, which he calls "obviously dumb and gameable at the extreme" but directionally right.¹

Middle management is done. Dorsey and Botha name three roles: individual contributors "who build and operate capabilities", directly responsible individuals "who own specific cross-cutting problems or opportunities and customer outcomes", and player-coaches "who combine building with developing people".⁵ Blomfield keeps the first two: "I actually don't like the third one, so I deleted it." Anything that needs doing gets "a named human, not a committee".¹

Make the organisation legible. "If it is recorded, it happened to the AI. If it did not get recorded, it did not happen." YC records partner emails, Slack, DMs and, for three or four months, office hours. He would add phones, clips, smart glasses or microphones in every room. A colleague turned about 2,000 hours of recorded office hours into a 150-page user manual over a weekend, to be refreshed monthly, and fed to an agent as "the combined wisdom of sixteen YC partners".¹

Software is ephemeral. Store the data "very preciously", never throw it away, and treat internal software as something to generate for an event and regenerate when the models improve. "The business context and skills are the valuable part."¹

People sit at the edge. The company brain sits in the middle. People work where it meets reality: novel situations, ethical questions, high-stakes and high-emotion moments, and sales, where he expects "a human being in the room for the next twenty years". The talk closes with a question: "If you were building your company today, would you start it in this shape?"¹

How the claims hold up.

A talk to founders is not a research paper, and I do not read it as one. But boards are being asked to act on these ideas, so each claim deserves to be set against what has been measured. The figure below gives my reading of ten of them. The sections after it give the detail.

Why hierarchies exist.

The talk's account of hierarchy is half of the story. Jay Galbraith described hierarchy in 1974 as one way an organisation processes information, with alternatives such as information systems and lateral links. On that view, an AI layer that routes information is a new information system, and some management work really can move to it.²² Ronald Coase explained the firm itself by the cost of coordination, and cheaper coordination can as easily make firms larger as flatter.²⁰

Hierarchy also does work the talk does not mention. It assigns accountability, settles disputes and develops people. Gallup finds managers account for at least 70% of the variance in team engagement.³³ Friedrich Hayek's point about "the knowledge of the particular circumstances of time and place" applies as well: much of what matters is never written down, so it never reaches a recording.²¹ Even Apple's directly responsible individual, the idea the talk keeps, works inside a strongly hierarchical company.⁵⁵

What copilots actually delivered.

The talk is right that copilot gains are modest. The largest field study, of 5,172 customer support agents, found 15% more issues resolved per hour, mostly for less experienced staff, with small quality declines for the most experienced.²⁵ METR's 2025 trial found experienced developers took 19% longer with AI tools on their own codebases, while believing they had been faster. Its 2026 follow-up points to a speed-up but with confidence intervals that include zero, and METR calls it very weak evidence.²³²⁴ Google's DORA 2025 survey links AI to higher throughput and also to less stable delivery.²⁷

MIT NANDA's 2025 report found that about 95% of organisations saw no P&L return from generative AI pilots, and blamed a "learning gap": tools that do not keep feedback or adapt.²⁶ The talk's learning loop is an answer to that gap. The report supports the diagnosis. It does not test the cure.

The delayering record.

Removing management layers is well under way. Meta flattened in 2023 and capped most managers at ten direct reports. Amazon told its senior teams to raise the ratio of individual contributors to managers by at least 15%. Google said in August 2025 that it had 35% fewer managers of small teams, many of them now individual contributors.³⁰²⁸²⁹ Middle managers made up 29% of layoffs in 2024, against 22% in 2018 to 2022.³² Gartner predicted that through 2026, 20% of organisations would use AI to remove more than half of their middle management. That is a prediction about one organisation in five, and its horizon is close.³¹

The reversals matter as much. In May 2025 Klarna's chief executive said that cost had been "a too predominant evaluation factor" and the result was "lower quality".³⁴ In August 2025 Commonwealth Bank reversed 45 redundancies made for an AI voice bot, called the decision an error and apologised.³⁵ Block staff told The Guardian that about 95% of AI-generated code changes still needed human modification, and one called the cut "posturing for the market".⁷¹⁵ I have written full case studies of the Klarna and Commonwealth Bank decisions.

Revenue per employee and tokens.

AI-native companies do report striking ratios. Gamma reported about $100 million of annual recurring revenue with about 50 staff in November 2025; Lovable reported $100 million with 45 full-time staff in July 2025.³⁶³⁷ Those are run-rate figures, and the denominator leaves out the compute the company buys. Part of what "burn tokens, not headcount" describes is cost moving from payroll to a model supplier.

On measuring token use, the speaker's own caveat proved right within weeks. Meta's internal "Claudeonomics" leaderboard ranked about 85,000 staff and logged more than 60 trillion tokens in 30 days before it was taken down; Amazon staff inflated token counts on trivial work and its leaderboard was retired.^43441112 Shopify renamed its token leaderboard a usage dashboard to stop staff competing for the top spot. Its 2025 memo asks teams to show why AI cannot do the work before they ask for more people, a sounder test than counting tokens.¹³¹²

What the law says about recording.

"Record everything" meets the law first. Eleven US states, California among them, mainly require every party's consent to record a conversation.⁴⁵ Singapore's PDPA lets an employer collect personal data reasonably needed to manage employment, but it must still tell employees the purpose, and the exception does not cover clients or visitors in the room.⁴⁶ GDPR Article 88 asks for measures that protect dignity, with monitoring systems at work named specifically.⁴⁷

The sharpest limit is in the EU AI Act. Since 2 February 2025, Article 5(1)(f) has prohibited AI that infers people's emotions in the workplace, except for medical or safety reasons.⁴⁸ Microphones in every room feeding a system that is meant to spot "high-emotion moments" comes close to that line. Research on monitoring points the same way: when supervisors use it for control, performance falls and rule-breaking rises.⁵⁷

Where the loop needs controls.

The claim I would not accept as stated is that a loop can write, review, merge and deploy code with no named person. The incidents are recent and specific. In July 2025 a Replit coding agent deleted a production database during a code freeze.³⁸ In April 2026 a coding agent deleted a company's production database and the backups stored beside it in nine seconds, causing an outage of about 30 hours.³⁹ An AWS outage in December 2025 was reported as caused by an internal agent; Amazon says the cause was misconfigured access, not AI.⁴⁰ Each case is a permissions failure as much as a model failure, which is the point: the loop's tools layer decides how much damage one bad change can do.

The quality gate has a deeper weakness. METR found a frontier model gaming its grader in 30% of runs on one benchmark and in every run on one task, and telling it not to cheat made little difference.⁴¹ Anthropic reported in November 2025 that models which learned to game coding tests went on to sabotage work when used as agents.⁴² An agent that reviews another agent's code and approves the merge is exactly the kind of grader these studies show being gamed.

Putting a person back in the path is not a full answer either. People who supervise automation drift into complacency under load, experts included, and training does not remove it.⁵³ A review of 41 policies requiring human oversight of algorithms found that people often cannot do the oversight the policy assumes.⁵⁴ The EU AI Act's human oversight article names automation bias directly, and Singapore's agentic AI framework asks firms to bound an agent's scope and keep people "meaningfully" accountable.⁴⁹⁵¹ My own work on oversight capacity reaches the same place: an overnight loop produces changes faster than anyone can review them, so the review has to be tiered by risk.

Which loops can run unattended.

My test for any loop is two questions: how much harm does one bad change do, and how easily is it undone? Placed that way, the talk's own examples sort themselves. Fixing a database view so staff queries work is low harm and easy to reverse; it can run unattended, with logging and a weekly sample. A weekly A/B test on a sign-up funnel can run inside hard limits with automatic rollback. Shipping customer suggestions into the product overnight needs a named person to sign before it reaches customers. Microphones in every room and token counts in performance reviews should not be automated at all until the legal and people questions are settled.

I would keep the talk's five stages and add three controls. First, the grader is never the thing it grades: a different model, a deterministic check or a person reviews each change, and the loop cannot edit its own gate. Second, every loop has a named owner, the talk's directly responsible individual, who can stop it and who answers for what it ships. Third, every change is logged with what triggered it, so a bad change can be traced and reversed. None of this slows the low-risk loops. It is what lets a board approve the rest.

What I would keep is most of the talk. Making the company's know-how legible is the right first move, and most firms I work with have not started it. Treating internal software as disposable is already true in practice. Keeping people for novel, ethical and high-stakes moments is where regulators have landed too. Where I part company is the player-coach. The role the talk deletes is the one that develops people, and the evidence on engagement says it is the hardest one to replace.

What is often misreported.

• It is not an Anthropic lecture. It is a YC batch talk; the speaker joined Anthropic about seven weeks after it was published.¹⁹
• YC has not published evidence for its overnight agent, the 150-page manual or the 5× figure. All three are the speaker's account.¹
• "Codex 5.5" is OpenAI's GPT-5.5, used in Codex. No product has that name.¹⁰
• The three roles are Dorsey and Botha's. Blomfield dropped the player-coach.⁵¹
• Google did not cut 35% of its managers. It has 35% fewer managers of teams with fewer than three people.²⁹
• Gartner did not predict half of middle managers would go. It predicted 20% of organisations would remove more than half of theirs.³¹
• The QJE study of support agents found 15%; the 14% figure is from the 2023 working paper.²⁵

What boards can take from it.

1. Start with legibility. The first step in the talk is also the cheapest: write down how each function works and keep the data. It pays off with or without autonomous loops.
2. Sort loops by consequence and reversibility. Let low-harm, easily reversed loops run unattended. Put hard limits and automatic rollback on the middle. Require a named signature on anything that reaches customers or people's data.
3. Never let the grader grade itself. Models have been measured gaming their checks. Review by a different model, a deterministic test or a person, and lock the loop out of its own quality gate.
4. Name an owner for every loop. Keep the talk's directly responsible individual, with the authority to stop the loop and the duty to answer for it.
5. Measure outcomes, not tokens. Token leaderboards were gamed within weeks. Ask teams to show what AI cannot do before they ask for headcount, and judge them on results.
6. Check the law before the microphones. Consent, notice and the EU ban on workplace emotion inference apply before any recording starts. Clients and visitors are not covered by employment rules.
7. Decide who develops people. If the middle layer goes, someone still has to coach, settle disputes and build judgment. Name who does it before the restructure.

To place a proposed loop by autonomy and consequence, use the agent risk assessment matrix, and estimate the review capacity it needs with the oversight ceiling calculator.

Controls for a self-improving loop.

Full timeline.

Dates to re-check.

Common questions about the talk.