Skip to main content
The factory is the product. Agents are outputs. That sentence inverts how the industry works. The default way to build an AI agent is artisanal: assemble one, tune it by hand until it behaves, ship it, start the next one from scratch. Every unit is a one-off; nothing compounds. This factory takes the opposite position: the durable asset is the production system — the stations, the gates, the standard work, the verification machinery — and every unit that moves through the line leaves the line slightly better than it found it.

​
What the factory actually manufactures

Follow the thesis one level down and it sharpens. Assembly is nearly free — the pilot line assembled a working unit in 49 seconds for $0.37 — and infinitely parallel. What is scarce is knowing the unit works. So what the factory manufactures is not agents. It is verified claims about agents — trust, in certifiable units. The agent without its certificate is inventory, not product. The plant’s Coda compresses it to one wall-sentence: Toyota’s plant takes steel in and puts trust out — nobody buys a Corolla for the stamping tolerances; they buy it because it starts every morning for twenty years. This factory is the same plant with the metaphor removed: skills, tools, and agents go in one end, and what comes out the other is a claim you can stake the operation on, with the evidence attached and an expiry date printed on it.

​
What the thesis buys

  • Owning the line beats owning any product off it. Products decay (certificates expire); the system that re-certifies them compounds.
  • Investment flows to the system first: verification before headcount, standard work before output, culture compiled into gates before culture written on posters.
  • The first customer is the factory itself. The line’s own tooling — its intake linter, its watchdogs, its gates — is built as units on the line, which is how a production system debugs itself.
The agents are proof the factory works. The certificates are proof you can sleep.

​
Self-audit