{"slug": "yc-described-the-self-improving-company-it-runs-on-a-brain", "title": "YC described the self-improving company. It runs on a brain.", "summary": "Y Combinator has added 'Company Brain' to its Requests for Startups, urging founders to build AI systems that capture a company's operational knowledge in plain, model-agnostic text so it can be continuously updated and acted upon by AI agents. Tom Blomfield, a YC group partner who has taken leave to join Anthropic, outlined the concept in a 13-minute talk to a YC batch, describing five recursive loops—sensors, policy, tools, quality gates, and learning—that make a company self-improving. Blomfield's example cites YC using thousands of hours of recorded office hours to regenerate its internal user manual, which now updates itself as new advice is given.", "body_md": "# YC described the self-improving company. It runs on a brain.\n\nTom Blomfield gave a thirteen-minute talk to a Y Combinator batch this spring, [How to Build a Self-Improving Company with AI](https://www.ycombinator.com/library/Qf-how-to-build-a-self-improving-company-with-ai) ([video](https://www.youtube.com/watch?v=X_JsIHUfUjc)). He then went further and put it in YC’s Requests for Startups, where “Company Brain” is now one of the ideas YC is explicitly asking founders to build. He has since [taken leave from YC to join Anthropic](https://www.pymnts.com/personnel/2026/anthropic-recruits-monzo-co-founder-to-tackle-compute-issues/), but the talk is still up, and it’s worth the thirteen minutes. Watching it, we kept having the same reaction. He is describing, from the outside and in his own words, the sort of company we built this+that to make possible.\n\n## Don’t bolt AI onto the old company\n\nBlomfield’s first move is to reject the framing most companies are using. Adding a copilot here and an assistant there’s, in his words, like taking the old way of working and putting a more powerful engine on it. The machine underneath is the same. What he tells founders to build instead is a set of recursive loops that improve themselves as they run, so the company gets better while you sleep.\n\nHe gives the loop five layers. Sensors, the real-world signal the company runs on: emails, support tickets, product telemetry. A policy layer that decides what the AI does on its own and what a person has to sign off. Tools, the APIs and actions the loop can actually take. Quality gates, the evals and human review that catch bad output before it ships. And a learning layer, where failures feed back so the next pass is better than the last. Run that consistently and the loop tightens on its own over time.\n\n## Make the company legible to AI\n\nThe instruction that stuck with us is the one underneath all five layers. For any loop to work, AI has to be able to read how the company actually operates, so Blomfield tells founders to capture it: the emails and chat where decisions really get made, written down somewhere a model can work from. He calls this making the company legible to AI.\n\nHis example is YC’s own. The firm took thousands of hours of its recorded office hours and used AI to regenerate its internal user manual, and that document now keeps updating itself as new advice is given. Those recordings are his example, not a template to copy; the point that generalizes is what he says is worth keeping. The precious thing is the understanding people have of how the work is done. The software sitting on top of it’s ephemeral, and if you store the knowledge as plain, model-agnostic text, it outlives whatever app happens to be reading it this year.\n\n## That legible knowledge layer is the brain\n\nWe have been making a version of this argument for a while, from a different starting point. A company’s real, live state, [what just slipped and who owes what](/blog/company-world-model-is-built-from-communication), lives in the messages its people send, not in the systems of record. The job is to turn that stream into knowledge an agent can read and, eventually, [write back to](/blog/agents-need-a-brain). We call that layer the brain.\n\nThe RFS goes further than the talk about what this has to be, and it focuses on the hard part. He asks for something that pulls knowledge out of every fragmented source, structures it, **keeps it current**, and turns it into an executable skills file agents can act on. Not a search tool. Not a chatbot over documents. His examples are the questions no schema has a field for: how your company handles refunds, the rules for pricing exceptions, how engineers respond to incidents.\n\nKeeping it current is the clause that has killed this idea before. Hand-authored ontologies died on upkeep rather than on formalism, which is an argument we have [made at length](/blog/semantic-web-failed-on-upkeep). Asking for a knowledge layer is easy. Asking for one that doesn’t go stale is asking for a solution to a problem no one has solved.\n\nThe brain is the legible company Blomfield is asking founders to build, packaged as a product. It holds the operational knowledge a team generates as it works, captured as text, kept current, and readable by any model. His office-hour example is meeting audio, and we don’t record or transcribe meetings. Our raw material is the written message stream across email, Slack, and Teams, plus whatever transcripts already land in it from wherever a team takes its notes.\n\n## His loop is our workflows and the brain\n\nHis five-layer loop maps, layer for layer, onto two of our marquee features.\n\nThe sensors are the messages that arrive. The policy and decision layer is the branching logic a workflow runs, the IF and the ELSE and the rules you set for what happens without you. The tools are the actions a workflow can take and the outside services it reaches through MCP. The quality gate is the approval step, the point where a run pauses and waits for a person before anything goes out. And the learning layer is write-back, where what a run produces returns to the brain so the next one starts from more than the last.\n\nWrite-back works today: a workflow can update the brain, so a team can keep its operational knowledge current as work happens, and we have built exactly that for a customer. What we haven’t shipped is the general-purpose version, a brain that maintains itself across the whole message stream with no one wiring it up. That’s the frontier, and it’s not what Blomfield’s team built either. Their example was a purpose-built loop that regenerated a single document from their own recordings, not a general self-maintaining brain. That version is ahead of all of us, and it’s what we’re working on with a design partner.\n\nWe are no longer alone in saying so. Since the RFS went out, a field has formed: **Hyper**, a YC Spring 2026 company calling itself the self-driving company brain, and others chasing the same brief. Garry Tan, YC’s own chief executive, wrote a company brain for himself and open-sourced it. Elsewhere in the category, Lore and Sana sell an “enterprise brain”, Glean sells enterprise search, and Guru sells knowledge management, all circling the same problem from different words. Nobody has shipped a knowledge layer that maintains itself across a company’s whole message stream.\n\n## The destination, and the on-ramp\n\nBlomfield is talking to founders who can staff this. His most vivid example is an agent that noticed database queries failing overnight, diagnosed the cause, wrote a fix, opened a pull request against YC’s codebase, had a second agent review it, and shipped it before morning. That result is impressive, and it assumes a team that can wire agents into its own code and infrastructure.\n\nMost teams cannot, and that is the opening. this+that can already hand a coding task to an agent that opens a draft pull request for a person to review, which is one rung on the same ladder. We could run this untended, too. The approval gate is optional and nothing forces a person to sign off on the pull request, so a team can wire up the same hands-off loop. We keep a person on it by default because we think that is wiser, not because the product requires it. Our product’s whole reason for being is that a legible, self-improving company should not require you to assemble the plumbing first. The knowledge layer assembles from the inbox you already have, and the loops run on top of it.\n\nWhen companies do reorganize around loops like the ones Blomfield describes, most of them will not get there by hiring a team to build the loop. They will get there because the knowledge was captured as they worked, and agents were already acting on it. Those are the companies we are building the brain for.\n\n## Key takeaways\n\n- In a YC batch talk, Tom Blomfield argues founders should stop bolting AI onto the old org and rebuild the work as recursive, self-improving loops, and lays out a five-layer version: sensors, policy, tools, quality gates, and learning.\n- His core instruction is to make the company legible to AI by capturing what its people know as durable, model-agnostic text, because the knowledge is the precious asset and the software on top is ephemeral.\n- That legible, self-updating knowledge layer is what we call the brain. We build it from the written message stream, not from meeting recordings, which is his example and not something we do.\n- His self-improving loop maps onto our workflows plus the brain: triggers as sensors, branching and rules as policy, actions and MCP as tools, approval steps as quality gates, and write-back as the beginning of learning. Write-back to the brain works today, so a team can keep its operational knowledge current now; the general-purpose, self-maintaining version is the frontier, and not what YC built either.\n- He is describing a destination that takes real engineering to reach. this+that is how a normal team gets there without building the loop from scratch, starting from the inbox it already has.", "url": "https://wpnews.pro/news/yc-described-the-self-improving-company-it-runs-on-a-brain", "canonical_source": "https://www.thisandthat.chat/blog/self-improving-company-needs-a-brain/", "published_at": "2026-09-02 10:22:26.813315+00:00", "updated_at": "2026-09-02 10:22:28.665472+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-products"], "entities": ["Y Combinator", "Tom Blomfield", "Anthropic", "YC's Requests for Startups"], "alternates": {"html": "https://wpnews.pro/news/yc-described-the-self-improving-company-it-runs-on-a-brain", "markdown": "https://wpnews.pro/news/yc-described-the-self-improving-company-it-runs-on-a-brain.md", "text": "https://wpnews.pro/news/yc-described-the-self-improving-company-it-runs-on-a-brain.txt", "jsonld": "https://wpnews.pro/news/yc-described-the-self-improving-company-it-runs-on-a-brain.jsonld"}}