The self-operating company

Most of what people do inside a company is move what they know to someone who needs it. Meetings, updates, handoffs, dashboards. By the time a company is a few hundred people, a great deal of its best attention is going into keeping the rest of it informed.

We would like that attention back, and pointed at the work that actually needed a person: judgment, taste, invention, and the relationships a business is made of.

All that carrying exists for one reason. The company itself does not know what it knows. Its decisions, its promises, and the dependencies between them are real, but they live scattered across the systems where work happens and the memories of the people who did it. Any two pieces can disagree, and nobody finds out until the disagreement costs something.

So the first thing a company needs is an account of itself. What has been decided, what has been promised, and what each of those things rests on, held in one place and kept current. When one part changes, the company should know what else moved. Almost everything else follows from that.

This is where most attempts will stop. Software engineering is the one place where agents genuinely work today, and that has less to do with the models than with the conditions: code has a single source of truth, and a cheap way to check whether an attempt was correct.

A company has neither. Its context is spread thin, and most consequential decisions are not right or wrong in a way you can test. They depend on history, on relationships, on what this company has chosen before. Building that sense of what correct looks like, here, in this company, is the problem worth working on, and it is the one we have chosen.

It would be easy to read all of this as a productivity gain, which is how AI is mostly being adopted: the same work, done faster, by fewer people. We think that badly underestimates what is happening. When the constraint that shaped an institution disappears, the institution does not get faster at what it was doing. It becomes something else.

Every company alive today was shaped by the attention of the people running it. The departments, the handoffs, the meetings, the whole architecture of modern work, is an accommodation to a limit we never had a choice about. We are the first generation who gets to ask what a company looks like without it.

We think the answer is companies that attempt things no organization built out of handoffs could hold together, run by people who spend their days on the parts that were always theirs. That is worth building carefully, and it is worth building now.

Chandra Bhagavatula has spent his career on the technical problems underneath this: memory, retrieval, and getting agents to run reliably in domains where being wrong is expensive. At Chip Stack he built agent harnesses for chip design. Cadence acquired the company, and that work became an AI super-agent now used at NVIDIA. The reliability pattern he built there is the one this needs.

Martín Ramírez is a computer engineer who spent the second half of his career on the business side of software. He was Chief Revenue Officer at WellSaid Labs, where he built the revenue engine from scratch. He is a two-time founder, and most of what he knows about operational debt he learned by accruing it.

The company is incubated at the AI House in Seattle, Washington. Both founders have spent years watching coordination cost slow down the companies they helped build, which is why they are working on this one.