An agency sells hours and slides. It staffs juniors against a senior’s name, rents the current model, and invoices as if attention were scarce. That business dies twice: first when the client learns to prompt, then when the model is cheaper than the junior. A factory sells a repeatable way to stand up a firm. The domain expert brings the field. The house keeps the machine. The thin company in the middle is allowed to be small. The intellectual property that makes the next one faster does not leave with the last invoice.
We are not interested in becoming a nicer agency. The world does not have a shortage of people who will “help you with AI.” It has a shortage of houses that can industrialise new-firm creation without pretending each one is a unique private-equity romance. Stripe’s public tell — more firms, faster — is a factory problem. You cannot partner your way through a ten-times rise in formation by adding another strategist.
The unit of work
The unit is not a campaign. The unit is: a person who already holds the license, the customers, or the plant; a once-off build that encodes how that work should run in software; a company thin enough to live or die without dragging the house; and a factory layer that stays here so the second and fifth times are not billed as the first.
That sentence is easy to corrupt. The corruption looks like this: the house becomes the agency (“we will build whatever you say”), the expert becomes a mascot, the thin company becomes a graveyard of prototypes, and the factory IP is accidentally left in a contractor’s laptop. The discipline is the opposite. The expert is not a vibe. They are the reason the software is about something. The build is once-off in the sense that we do not reopen the architecture for taste. The company is thin so that failure is local. The factory stays in the house so that success is not a one-time fee.
Give operators more economic control — equity and product — rather than a body shop of hours the model deletes. That is the second Stripe goal, translated out of a payments letter and into a fiduciary habit. If the only thing the expert receives is a service agreement, we have rebuilt 2019 with better typography.
What a factory refuses
It refuses the retainer that exists to keep people busy. It refuses the workshop whose output is a deck about workshops. It refuses the white-label of a public chatbot with a family colour on the button. It refuses to staff a room of graduates to click through a model the house does not run. Those are agency motions. They feel like work. They do not accumulate an asset.
It also refuses the opposite vice: treating every operator as a founder-messiah who must raise a round. Most licensed trades do not need a round. They need a successor, a cleaner set of books, and software that does not insult the people who still go to the site. The factory can serve that person without dragging them into the venture book’s mythology. If the result is a real company with cashflow, it may belong in the haven. If the result is a repeated motion that spawns new firms, it belongs in the book that may die. The factory is a method. It is not a license to mix the books.
Industrialising formation without a ticket machine
Private equity trained a generation to think that the only grown-up way to do something twice is to buy it with a large ticket and a story about synergies. That is one method. It does not survive a world in which new firms appear weekly and the scarce input is no longer a junior associate. A factory is the other method: encode the motion, keep the motion, let the vehicle stay thin.
This is not a claim that software replaces the expert. It is the opposite. Software without the expert is an agency demo. The expert without software is a practice that cannot be inherited except by another expert who may not exist. Together they are a firm that a model can wrap rather than replace. Wrap means: the chatbot does the middle office; the licensed human still signs, still climbs, still holds the ticket from the state.
When we say the factory stays with the house, we mean the repeatable layer — prompts are the least of it; the workflow, the document spine, the settlement, the way a new entity is stood up without a new religion. The expert should get richer if the thing works. The house should get faster. Nobody should get a surprise invoice for the fifth copy of a wheel.
Agency is a recession and a singularity victim
In a downturn, clients cut the hours they never understood. In a phase change, they cut the hours a model now understands. An agency is short both proofs. A factory that owns its motion can still be wrong — products die — but it is not wrong in the same way. It has an asset that is not a timesheet.
We will not name the experts. We will not publish a shop window of “verticals.” A journal that lists trades as if they were inventory is a deal room that forgot to close the door. The public claim is the shape: domain expert, once-off build, thin company, factory in the house. If that sounds less exciting than a terminal animation, it is because factories are supposed to be dull on the third run. Dull on the third run is the point. The first run can be fun. Fun that cannot be repeated is an agency offsite.
The house will still sit next to people who want to change the world. That appetite belongs in the venture book. The factory is how we meet them without hiring a new firm of billable hours every time the calendar says another company was born. Formation is the tell. We intend to be on the side that industrialises it — and to stay a house, not a shop.