The Company Brain: The New Physics of Business

The new physics of business

At the AI Engineer conference, Garry Tan — president of Y Combinator — opened with a number designed to make people uncomfortable: his personal productivity has multiplied by 400 since 2013. He is the first to discount it: apply the harshest penalty you like for bloated code or self-flattery, and it is still an enormous leap. But the real message is not about numbers. It is about physics: the rules of how much one person can accomplish in the same hours have just changed.

The most valuable part of the talk is the explanation. The people getting 2x and the people getting 100x are using exactly the same model, the same weights, the same API. The leverage is not in the AI; it is in how you wire the work around it.

And the most revealing part is who is doing that wiring. At YC itself it is not just the engineers: it is the media team, the events team, the finance team. People who had never opened a terminal are building skills and automations; one finance person turned a hundred spreadsheets into an application she built and maintains herself. In Tan's words, she has not become a programmer: she has become a manager of agents.

That is the reading we care about. The new physics is not about needing fewer people: it is about every person being able to orchestrate their own team of agents, delegate the repetitive work and spend their time on what genuinely requires human judgment — thinking more strategically, making better decisions, staying closer to the customer. And for that to work across an organization you need something almost no company has yet: a company brain.

Your company is a library (and your agents can only open three books)

Tan uses a vivid image. A human being holds about seven things at once in working memory; that is why phone numbers have seven digits and why we forget the eighth item on the grocery list. Every checklist, every org chart and every filing cabinet humanity has invented is a prosthetic for that limit.

An AI agent, on the other hand, holds around a million tokens of context: about a thousand pages, or three Harry Potter books open at once, in which it can find a needle and synthesize across all three in seconds. That is an enormous leap — and at the same time very little. Because a company is not three books: it is an entire library. Every email, every meeting, every decision with its reasoning, every customer conversation, every closed project.

The question that decides whether your agents behave like geniuses or like goldfish is who chooses which three books are open on the desk at any given moment. That discipline is what we now call context engineering, and the system that solves it is what Tan calls the company brain: the library plus the librarian.

Why it is not "just RAG"

The first reaction is usually that this is plain old RAG: retrieve the relevant documents and hand them to the model. Tan's answer is yes — in the same way a database is "just" B-trees. Retrieval is the primitive; the hard part is everything around it: what gets written down in the first place, how it gets enriched and linked, what is promoted to hot memory and what is filed as cold reference, and who arbitrates when two facts disagree. His line sums it up: retrieval is easy; being worth retrieving from is the product.

He also warns about the failure modes, which anyone who has tried this will recognize. A brain nobody curates becomes a garbage dump with excellent search. Retrieval will surface a stale fact with total confidence. A bad procedure encoded as a skill perpetuates a bad process forever. Tan's conclusion is that the primitive is not memory but memory plus hygiene: provenance on every fact, contradiction checks, and a librarian — human plus agent — whose actual job is pruning.

Put another way: the company brain is not a technology problem, it is a management problem. And that is exactly where Minte works.

How we make it real at Minte

Everything Garry describes with markdown files and personal discipline needs a product layer inside a company: multi-user, with permissions, with traceability, and designed for everyone to use — not just the technical team. These are the pieces Minte uses to turn the speech into something you can actually manage.

The graph: a map of your organization

A brain is not a drawer full of documents; it is a network. Minte's graph connects people to their teams, teams to their objectives, and objectives to the documents, conversations and customers around them. When an agent answers, it does not start from a blind search box: it starts from a map of how your organization fits together. It is the difference between asking a brand-new intern and asking someone who has been in the company for years.

The librarian: RAG with judgment and permissions

Retrieval works on top of that map. Minte combines exact-match and semantic search to find the right passage even when nobody remembers the document's title. But the important part happens before searching: each workspace decides which knowledge folders every portal, team and role can see. Minte's librarian only opens the books that person is allowed to read, and every answer arrives with its sources, so you can always trace where a fact came from.

Skills: your team's know-how, turned into procedure

Tan puts it plainly: a skill is one capability, one job, written clearly enough for someone to execute. From there comes his golden rule: never do one-off work; if you have to ask for something twice, you failed. In Minte, skills are born from the conversation itself: when you like a result, it becomes a saved procedure any colleague can repeat tomorrow. And it is the teams themselves — sales, operations, finance — who create and maintain their skills, without waiting on a technical department. The organization that captures what it learns gets smarter every single day; the one that does not wakes up every morning with amnesia.

MCP: the agent's hands

A brain without hands is just an observer. MCP (Model Context Protocol) is the open standard that connects agents to the company's real tools: the CRM, the ERP, calendars, databases. Minte aggregates multiple MCP servers in a single workspace and controls which tools each agent can use, with intelligent filtering so a giant API response does not drown the context — remember, only three books fit on the desk. And with microapps, a connected process becomes a small application any team can use without ever opening a terminal.

Governance: what turns the experiment into infrastructure

This is the difference between the lab and the enterprise. In Minte every company lives isolated from the others, permissions work on two levels — functional and organizational — sensitive actions ask for human confirmation before executing, and everything is logged: what each agent did, with what context, and at whose request. It is the practical translation of Tan's advice: treat your brain like production infrastructure and it compounds; treat it like a dumping ground and you get a very confident agent being wrong in ways nobody can trace.

Start with one book, not the whole library

Tan's personal brain holds 220,000 pages, mostly written by his own agents. Nobody starts there. You start with one well-curated knowledge folder, one process turned into a skill, one integration connected through MCP. Compounding does the rest: every linked document, every captured procedure and every verified source makes the next one easier.

The goal is not to replace anyone: it is for every person on your team to hand the mechanical work to their agents and win back time to think, decide and create. Tan closes with a line worth keeping: model quality is rented; your brain is yours. Models will keep improving for everyone equally. The library, the librarian and the procedures your team builds this year are the part nobody can rent. If you want to see what your company's brain would look like — and how your team can become its best librarian — let's talk.