The Company That Fits in One Repository

There is a company in Vilnius whose entire body — its shopfront, its filing cabinet of half a million companies, its salesman, its back office, and the keys to its own building — fits inside a single folder on one laptop.

You could read the whole of it in an afternoon.

The folder is what programmers call a repository, or repo: think of it as the master ledger of a business, except instead of accounts it holds every instruction the business runs on.

Northwestern Solutions, a Lithuanian UAB, keeps all of it in one such folder.

Inside are the frontend — the public website customers see, the equivalent of your shopfront and signage — and the backend, the machinery behind the counter that nobody sees: a database of more than 500,000 Lithuanian companies, the engine that drafts sales emails, the password-locked back office where the owner approves each one, and the deploy scripts that put the whole thing onto a rented server, the way you'd hand a foreman the keys and the opening procedure.

One hundred and twenty files. No department. No org chart.

It is built and run by one person.

Not by typing every line himself, but by directing agents — AI workers you assign a goal to, the way you'd brief a capable employee, who then plan the job, do it, check their own work, and hand it back for approval.

One human, a workshop full of tireless apprentices, and a single ledger they all read from.

I want to convince you that this is not a novelty. It is the early shape of something that should interest anyone who still measures a company by how many people it employs.

You have felt the ground move even if you couldn't name it. The signs were there before AI. Instagram sold to Facebook for a billion dollars with thirteen employees.

WhatsApp sold for nineteen billion with fifty-five — roughly $345 million of value per person on the payroll. Mojang, the maker of Minecraft, sold for two and a half billion with about forty.

The old law that value requires a proportional crowd of people was already bending.

What AI has done is turn a curve into a wager.

Sam Altman, who runs OpenAI, has said that among his circle of technology executives there is a literal betting pool on the first year someone builds a one-person, billion-dollar company — something, he said, that "would have been unimaginable without AI.

And now it will happen." Treat that as a prediction from an interested party, not a fact. The fact underneath it is duller and more convincing.

The fact is a number any capital-minded owner already respects: revenue per employee. The average publicly traded software company produces roughly a few hundred thousand dollars of revenue per head.

The new AI-native firms, by reporting compiled in Forbes and by the startup-data firm Dealroom, are running into the millions per head — Dealroom clocked the coding-tool maker Cursor above $3 million per employee.

That is not a better quarter. That is an order of magnitude — a break in the exchange rate between people and output. When a ratio your industry has lived by for a century moves by that much, it is not a trend.

It is closer to a change in the physics.

Here is the part that sounds like a technical detail and is actually the whole point.

An AI agent does its best work when it can see the entire job at once.

Split a company's code across several folders — a common practice, done for tidiness — and each boundary becomes what one engineer, Francis Eytan Dortort, calls a "context wall."

Put a wall between the shopfront and the back office and the agent, like a new hire who can only ever see one room, starts guessing about the rest.

Keep everything in one repository and, in his phrase, "the agent has one graph to work with."

It can trace a decision on the website all the way down to the database and back, against the real thing, not a stale memo describing it.

Two facts made this possible at once. First, the old proof that everything-in-one-place scales: Google keeps roughly two billion lines of code in a single repository used by tens of thousands of engineers.

That predates AI entirely. Second, the agents' field of view caught up with reality.

An AI's context window — how much it can hold in mind at one time, measured in tokens, which are word-fragments — went from about 200,000 tokens to a million in roughly a year.

A million tokens is enough to hold a small company's entire technical body in a single thought.

That is the quiet mechanism behind the one-person firm.

Not that the AI got smarter in the abstract, but that a whole business finally became small enough to fit inside one mind — and one folder is how you hand it over whole. (The honest caveat: advertised context and usable context differ; models can lose track of things buried in the middle of a very long document, which is exactly why the folder still needs to be well-organized.

The apprentice is tireless, not infallible.)

The rhythm is closer to running a workshop than to writing software.

You give an agent a goal — "add a page that shows each company's tax-debt history," say. It plans the work, edits the files, runs the tests that check nothing else broke, and hands back the result.

Anthropic's 2026 report on agentic coding describes cycle times collapsing "from weeks to hours," and documents a striking case at the Japanese firm Rakuten: an agent working inside a codebase of 12.5 million lines completed a complex task in about seven hours of unsupervised work, in a single run.

The human's job changes shape. You stop being the typist and become the foreman.

The investor Tomasz Tunguz, who tracks this closely, says he personally runs about four agents at once; the best operators manage ten to fifteen.

And — this is the number that keeps it honest — a large share of what the agents produce gets rejected and sent back. The work that remains for the human is not the doing.

It is the two things machines still cannot do for you: writing a clear brief, and judging whether what came back is any good.

Andrej Karpathy, one of the field's clearest voices, describes the craft moving from what he once playfully called "vibe coding" toward "agentic engineering" — the operator becoming an orchestrator of fallible, stochastic agents.

The value drains out of typing and pools in specification and verification.

Step back and this rhymes with something the investor Naval Ravikant said years ago. There are, he argued, two kinds of leverage.

The old kind needs permission: to command labor you must be hired or must hire; to command capital someone must entrust it to you. The new kind needs no one's permission.

An employee goes home. A machine that sends and drafts and files does not. That is the philosophical engine under all of it.

But the deeper move is not about leverage; it is about recomposition.

For two and a half centuries, since Adam Smith watched pins being made, the way to grow was to divide labor — to chop a job into ever-narrower slivers and hire a specialist for each, then build hierarchy to glue the slivers back together.

The assembly line, the department, the org chart, the ERP system stitching the silos: all of it is the machinery of that division.

Two commentators writing for the Swiss financial paper AGEFI, Comtesse and Eichenberger, argue that this Taylorist company — "division of labor, integrated management, siloed processes" — is now "collapsing."

What AI does is let the slivers recombine. The fragments the assembly line once split apart fold back together inside one accountable person.

Which returns us to a figure who predates the factory entirely: the master craftsman.

The one who held the whole enterprise in his head — who knew the customer, the material, the books, and the making, because he was all of them.

The Industrial Revolution displaced him by proving that ten narrow specialists out-produced one generalist.

The wager now is that leverage has swung back, and the generalist who can direct a workshop of machines out-produces the ten.

The scarce skill stops being how much you know or how fast you work; it becomes how well you can allocate and direct the resources that do the work. The owner-operator returns — with industrial horsepower this time.

Now the cold water, because you have earned it.

The most sobering finding in the whole field comes from a research group called METR, which ran a proper controlled trial.

Experienced developers using early-2025 AI tools were 19 percent slower than without them — while believing they had been sped up by 20 percent. Read that twice.

The tools made them feel faster and measurably weren't. The lesson is not that the tools are useless; it is that your gut is not a measuring instrument. Only the ledger tells the truth.

If you cannot count the result, you do not know the result.

Then there is the ceiling on what agents can actually do.

They are strong on well-defined, checkable work and weak exactly where your business lives: deep domain expertise and the undocumented logic that exists only in a veteran's head.

An NYU professor, J.P. Eggers, put the risk plainly: with an AI-built operation "you're kind of taking it on faith that what the AI is producing is pretty good."

The clearest cautionary tale is Base44 — a company one man, Maor Shlomo, built largely solo with AI agents, that made nearly $1.5 million in its first month and sold to Wix for $80 million within months.

A triumph for the thesis.

But the same reporting records that he set alarms every two or three hours to nurse the servers, shut off his own AI support bot after two weeks because he needed direct sight of the tickets, and ultimately recognized that scaling demanded expertise he did not have.

As he put it: "eventually… I need help." The machines bought him reach, not omniscience.

And the bills are real — always-on agents at lean startups can run into the hundreds of thousands of dollars a month, rivaling the salaries they replace.

When the same tools are available to everyone, advantage moves elsewhere — to taste, trust, distribution. When everyone is super, as the old line goes, no one is.

This is where a traditional executive's instincts are an asset, not a handicap.

Adoption has sprinted ahead of control.

The analyst firm Gartner expects 40 percent of enterprise software to embed AI agents by the end of 2026, up from under 5 percent a year earlier — yet only about one company in five has a mature way to govern them.

And the danger is subtler than a robot running amok.

Microsoft's security team, after a year of attacking these systems, found that a common weakness was not the AI but the oversight itself: a human "in the loop" who, worn down by a hundred approval clicks, starts rubber-stamping.

Presence is not practice. A signature given without reading is not a control; it is theater.

The safe pattern has a name — earned autonomy — and it will sound like ordinary prudence to you, because it is. Approve everything at first.

Expand an agent's freedom one category at a time, and only on a proven track record.

Match the strictness of approval to how reversible the action is and how big the blast radius: let it reformat a draft freely; make it beg permission before anything leaves the building.

Keep an unerasable log of who approved what.

Northwestern's outreach engine hard-codes exactly this. One module, and only one, is allowed to touch email.

Before any message can go out it must pass, in order: a suppression list of anyone who ever opted out (honored forever); a check that no company gets contacted twice in a campaign; a daily ceiling counted in plain arithmetic the AI cannot argue its way past; a mandatory legal footer; and, last, a human being clicking approve in the back office.

The AI may add a name to the do-not-contact list. It is structurally forbidden from removing one. That is not distrust of the machine. It is the discipline that lets you trust it at all.

So price a business differently. Stop valuing it by its payroll — headcount was always a proxy, and the proxy is breaking.

Start asking the two questions that increasingly separate an AI-native operation from a legacy one: how much revenue moves per accountable human, and how sound is that human's judgment. Those are the scarce inputs.

Much of the rest is getting cheap.

And notice what agents cannot touch. Your relationships, forty years deep. The trust a customer extends to a face, not a login.

Your distribution, your feel for the material, your standing with a regulator who knows your name. Those are not weaknesses to be automated away. They are the moat.

The move is not to defend against this technology; it is to bolt your irreplaceable assets onto its leverage.

Begin where a mistake costs nothing. One process. Human approval on everything. Autonomy earned, never assumed.

And — the discipline Northwestern binds itself to — do not rebuild your operation before a single paying customer has proven the loop works. Leverage applied to an unproven idea only helps you fail faster.

For two and a half centuries, to grow meant to add people, and to add people meant to add the coordination that binds them — the meetings, the memos, the managers of managers.

For the first time, growth can plausibly run the other way: similar output with a fraction of the org chart, because the coordination itself is what got subtracted.

The company that fits in one repository is not a smaller company.

It is a more legible one — a business a single person can once again read from end to end, and answer for, the way an owner could before the factory made that impossible.

That is the question worth sitting with. Not whether to hire fewer people; that is a cost decision, and a small one.

The real question is whether you could once more know your whole enterprise — every room of it, all at once — and stand behind all of it personally.

For a hundred years that was a thing only the smallest shopkeeper could say. It is becoming sayable again, at scale.

Whether that is a burden or a homecoming is, I suspect, the most revealing thing you could learn about yourself as an owner.