What Y Combinator Says About the Autonomous Company

Y Combinator publishes a list of ideas it wants founders to work on, called Requests for Startups. Each one is a short note from a partner, and the list changes with every batch. Read one at a time, they look like separate bets: legal work, insurance, compliance, internal tools. Read in order from 2025 to late 2026, they tell a single story about how companies will be built and run once AI can do real work.

This article pulls that story together from YC's own public material: the Requests for Startups pages, and public remarks by YC's president and CEO Garry Tan. It explains what YC's partners actually say about AI-native companies, where their picture matches the idea of the autonomous company, where it leaves gaps, and what an operator outside the startup world can take from it this quarter.

What does Y Combinator mean by an autonomous company?

An autonomous company, as this site uses the term, is a company that notices what needs doing, decides, delegates and checks its own work, while people keep the final say. YC's partners describe the same shape in their own words, spread across several requests.

The clearest single statement comes from Diana Hu's Summer 2026 request, "The AI Operating System for Companies". She writes that the best AI-native companies YC sees have "made their entire company queryable", and that this "turns a company from an open loop into a closed loop." In her open loop, "you make a decision and maybe check the results weeks later." In the closed loop, "the system monitors what's happening, compares it to what should be happening, and adjusts."

That is close to a working definition of autonomy. A company runs itself to the degree that it can see its own state, compare it with its goals and act on the difference. YC's other requests add the parts around it: who does the work, what they need to know and how many people sit on top.

Three terms recur in YC's writing and are worth separating:

• AI-native: built around AI from the first day, so the workflow assumes a model can read, write and act.
• Full-stack AI company: a company that uses AI to deliver the end service itself, instead of selling AI tools to the firms that deliver it.
• Closed loop: a company that checks its results against its intentions continuously, instead of at a monthly review.

The autonomous company combines all three, and adds the human side: clear owners, approval rules and proof that work happened.

How did YC's Requests for Startups move from tools to doing the work?

The direction is easiest to see as a timeline. Each row below is a request on YC's public page, with the season it was listed under.

Gustaf Alströmer's note spells out the history in three steps: "Historically, services became SaaS software. More recently, they became AI copilots." The step he asks founders to take next is AI-native companies that "sell the service." His reason is plain arithmetic: "The total spend on services is many times larger than the spend on software."

Jared Friedman made the same move a year earlier with the law-firm example. His point was that most founders build an AI agent and sell it to law firms, while a full-stack founder builds the firm. The subject of the sentence changes. A tool helps someone at a company do work. A full-stack AI company is the company doing the work. This site covers the business-model side of that shift in service as software.

For an operator, the useful reading is about what is now considered normal. If YC treats "AI does the whole task" as the baseline for new companies, customers and investors will soon compare existing companies against it.

Why does YC think small teams can now run large companies?

Aaron Epstein's Fall 2025 request makes the case directly. He writes that new AI tools make it "now possible" for small, high-agency teams, including solo founders, to build multi-billion dollar companies "with as little as just $500k in funding from YC." He compares it to cloud computing, which removed the need to buy servers, and argues that small teams avoid "the politics, excessive meetings, and lack of focus that grinds huge companies to a halt."

Garry Tan gave numbers from inside YC. In a March 2025 interview with CNBC, he said that for about a quarter of the then-current batch, 95% of the code was written by AI, and that the Winter 2025 batch grew 10% per week in aggregate. "What that means for founders is that you don't need a team of 50 or 100 engineers," he said. CNBC also reported his remark that companies were reaching as much as $10 million in revenue with teams of fewer than 10 people.

Two cautions keep this grounded. First, both statements are about software startups, where code is a large share of the work. A manufacturer has physical steps, buyers and auditors that code cannot replace. Second, a small team does not mean a small amount of coordination. It means the coordination has to be done by something other than people in meetings. That is the gap the next two requests try to fill. The site's article on the one-person unicorn works through the arithmetic of very small teams in detail.

What does YC say is missing: the company brain and the closed loop

The two Summer 2026 requests from Tom Blomfield and Diana Hu are the most relevant to anyone running an existing company, because they describe infrastructure every company would need.

Blomfield's "Company Brain" starts from a claim worth repeating: "The biggest blocker to AI automation of companies is no longer the models, they just got so good so quickly. Now the blocker is the domain knowledge." He describes the problem every operator knows. Know-how lives "in people's heads" and in old email accounts, chat threads, support tickets and databases, and "the company works because humans vaguely remember where that knowledge is and how to apply it." His answer is a system that pulls knowledge out of those sources, keeps it current and turns it into "an executable skills file for AI." He adds that this "isn't a company-wide search or a chatbot over documents. It's a living map of how a company works." He also names Garry Tan's own G-Brain as the model. Tan gave a talk on the same theme, "Every company should have a Brain", at the AI Engineer World's Fair in 2026.

Hu's request goes one step further. Knowing how the company works is half of it. The other half is noticing, continuously, when what is happening differs from what should be happening. She says building this today "requires brutal integration work" across chat, issue trackers, code, documents and call recordings, and that there is no product that "connects all this context into a single intelligence layer that can reason across it, flag when engineering is building the wrong thing, or generate specs agents can execute on." Her closing line asks for "the system that turns a company's own artifacts into a self-improving loop," and she adds: "Not another dashboard."

  OPEN LOOP                         CLOSED LOOP
  decide ──▶ act                    sources ──▶ notice the gap
              │                                    │
         weeks pass                        compare with goals
              │                                    │
         someone asks                        owner + date
         "what happened?"                          │
                                           act ──▶ proof ──┐
                                            ▲                │
                                            └── learn ◀──────┘

Put together, the two requests describe two layers. The brain holds what the company knows. The closed loop decides what that knowledge means today. This site argues the same split in why a company brain is not enough: memory remembers information, and intelligence understands what it means for the business right now.

Where YC's picture of the autonomous company stops short

Requests for Startups are short by design. They point at a market, so they leave out the operating detail a company needs before it lets AI act on its behalf. Four gaps stand out.

• Ownership. A closed loop that flags a problem still needs one person who answers for it. A flag sent to a channel is read by everyone and owned by no one.
• Approval. "AI does the work" raises the question of which work it may do without a person saying yes. Payment terms, prices and anything sent to a customer usually need a human approval.
• Proof. A task is done when there is evidence: the email that went out, the meeting on the calendar, the report filed. An agent saying "done" is a claim.
• Judgment under pressure. Edge cases, unhappy customers and people trying to game the system are where autonomous setups break.

Anthropic's experiment is a useful check. In Project Vend phase two, an AI agent ran small shops with an AI "CEO" above it. Anthropic reported that the changes "did make Claudius's shop more successful," but that the agents "still needed a great deal of human support," and summed up one lesson as "we rediscovered that bureaucracy matters." Procedures, checklists and approval rules are part of what makes autonomy work.

Common mistakes when reading YC's requests

• Reading them as forecasts. They are ideas YC would like to fund. YC's own page says they "represent just a fraction of what we fund." Treat them as signals of direction.
• Copying the software numbers. A coding team that writes 95% of its code with AI is a different case from a factory. Look at which of your tasks are information work before setting targets.
• Building the brain and stopping. A store of company knowledge with no loop that acts on it becomes a better search box. Blomfield himself says the brain is the layer between data and "reliable AI automation", which implies something that acts.
• Skipping the human roles. Fewer people does not mean no people. It means each person owns more outcomes and approves more actions, so the rules for who decides what matter more.

How to start if you are not a YC startup

Most companies cannot rebuild themselves as AI-native from scratch. They can borrow the closed loop in stages.

1. Make one team legible.

   Pick the team where dropped work costs the most, often sales or operations. Connect its email, calendar and chat read-only, so the record of commitments exists in one place.

2. Write the goals down.

   The loop needs something to compare against. List this quarter's goals and the promises that serve them: orders, audits, launches, renewals.

3. Turn gaps into owned items.

   Each time what is happening differs from what should be happening, create one item with one owner and a date. Avoid alerts to groups.

4. Let agents prepare, and people approve.

   Drafts, follow-ups and reports can be prepared by AI. Hold anything outbound for a person's approval at first.

5. Close on proof and review monthly.

   Count what closed on evidence by its date. Widen what can run with lighter approval only where the record shows it holds up.

For the same direction seen from another accelerator, read what a16z speedrun founders are building.

How GeniOS approaches the autonomous company

GeniOS reads the places where a company's work happens: email, calendar, chat and the business tools it already uses, read-only to start. It forms the company's goals from its own emails, decks and meetings, and checks every team's pending work against them. In YC's terms, it makes the company legible and closes the loop.

What reaches a person is intelligence: one situation, what to do, why, one owner and a date. The owner answers Delegate, Done, Later or Not relevant. Delegate hands work to a person or an AI agent, and nothing is sent until a human approves. An item closes only on proof, such as the email that went out, and GeniOS learns from each outcome.

That gives the four parts the requests leave implicit: an owner, an approval, proof and learning. The aim is the autonomous company as the GeniOS manifesto describes it: a company that notices, decides, delegates and checks its own work, while people keep the final say.

Worked example: a closed loop at a safety-shoe maker

The names, dates and numbers in this example are made up to explain the idea. Vaultex makes safety footwear in Greater Noida and is GeniOS's first design partner. Its week has three large threads running at once.

• Bata wants 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000, with a vendor review on Friday. Bata expects 60-day payment terms.
• Reebok samples are running on Line 2.
• Adidas has a vendor audit coming and needs IS 15298 test reports.

In an open loop, each thread lives in someone's inbox. Harsh in sales knows Bata wants a revised quote. Anushka in production knows Line 2 is busy. Nobody has connected the supplier email saying steel toe caps are up 8% to the $12 price. The gaps surface on Friday morning.

In a closed loop, the system reads the same sources and compares them with what should be happening before the review:

An AI agent drafts the revised quote and the audit cover note. Harsh and Anushka approve before anything is sent. Each item closes when the evidence exists. By Friday the review starts from a finished picture. The Vaultex case study covers how this works in practice.

Conclusion

Y Combinator's partners have described the autonomous company in parts: AI that does whole tasks, firms that sell the finished service, tiny teams with large revenue, a company brain that holds how the business works, and a closed loop that notices when reality drifts from plan. Add owners, approval rules and proof, and the parts become a company that runs more of itself while people keep the final say. Any company can start that path with one team and a read-only connection. To see how it looks with your own data, see pricing or book a call.

Sources

• Y Combinator, Requests for Startups: The AI Operating System for Companies (Diana Hu), Summer 2026
• Y Combinator, Requests for Startups: Company Brain (Tom Blomfield), Summer 2026
• Y Combinator, Requests for Startups: AI-Native Service Companies (Gustaf Alströmer), Summer 2026
• Y Combinator, Requests for Startups: AI-Native Agencies (Aaron Epstein), Spring 2026
• Y Combinator, Requests for Startups: The First 10-person, $100B Company (Aaron Epstein), Fall 2025
• Y Combinator, Requests for Startups: Full-stack AI Companies (Jared Friedman) and Internal Agent Builder (Pete Koomen), Summer 2025
• Y Combinator, Requests for Startups: Vertical AI Agents (Jared Friedman), Spring 2025
• CNBC, Y Combinator startups are fastest growing, most profitable in fund history because of AI, 15 March 2025
• AI Engineer, Every company should have a Brain (Garry Tan), 2026
• Anthropic, Project Vend: Phase two, 18 December 2025