The line, station by station
Seven stops take a request from the outer wall to the loading dock. Each one has a plainer name on the floor than it has in the trade.
| On the floor | Real name | What it does |
| 01 | The order window | Intake, or triggers | Turns a ticket, a message, an alert or a schedule into work. |
| 02 | The blueprint desk | Specification | States what is wanted and how it will be checked. |
| 03 | The cell | Sandbox: a sealed, disposable workspace | Gives each job a sealed, disposable copy of what it needs. |
| 04 | The tool crib | Shared tools and company context | Issues your systems, documents and rules, under permissions you set. |
| 05 | Many cells | Parallel agents and orchestration | Runs many jobs at once and hands out the work. |
| 06 | The inspector and the test track | Verification | Checks the result independently, against tests the builder never saw. |
| 07 | The gate and the control room | Approval policy and monitoring | Decides what ships, and shows people how the line is doing. |
The dial, level by level
| Level | Name | Who does the work | People on the floor | Why we trust it |
| 1 | Hand tools | A person, with AI finishing their lines | 19 | A person did the work |
| 2 | Power tools | AI, one piece at a time, when asked | 19 | A person read every line |
| 3 | The line | AI, start to finish, unattended | 3 | A person approved it at the gate |
| 4 | The trusted line | AI. Routine work ships by itself | 2 | The tests passed, and a person samples |
| 5 | Lights out | AI | 0 | The tests passed |
The head counts describe this model floor of 16 cells. They are an illustration, not data.
Will it work beyond software? Four questions
Run these on one piece of your own work. The factory fits as far as the answers are yes.
- Is the request written down?If work starts with a ticket, a form or a message, it can become an order.
- Can a machine check the result?If only a person can tell whether it is right, a person stays at the gate.
- Can it be rehearsed safely?The work needs a sealed copy to practice on, away from real customers and real money.
- Can a mistake be undone?If it cannot, someone signs before it leaves.
| Order | Written down? | Machine can check? | Safe to rehearse? | Can be undone? | Where it ends up |
| Match a supplier’s invoice to the purchase order | Yes | Yes. The numbers agree or they do not. | Yes, on a copy of the books | Yes, until it is paid | Runs through. People audit samples. |
| Answer a customer asking “where is my parcel?” | Yes | Partly. Facts yes, tone no. | Yes, as a draft | No, once it is sent | Stops at the gate. A person signs the unusual ones. |
| Decide next year’s prices | Rarely | No | No | Not easily | Lifted off the line. AI helps a person think. |
The factory is working in software at a handful of teams, and early everywhere else. These three ratings are my own illustrations to make the test easy to follow. They are not measurements.
Terms that get mixed up
- Agent
- An AI model allowed to work in rounds: act, look at the result, act again, until the job is done.
- Background agent
- An agent that works by itself in a remote workspace, started by a ticket or a message, with nobody watching it work.
- Harness
- Everything built around the model so it can do the job: the tools, rules, tests and feedback it is given. Most of a factory is harness.
- Orchestration
- Handing out work to many agents at once and keeping them out of each other’s way. The dispatcher board.
- Software factory
- The whole line: order window, blueprint, cells, tool crib, inspection, gate and control room.
- Dark factory
- A factory with no people on the floor. The phrase comes from manufacturing, where a plant run by machines needs no lights.
Who is doing this in public
Each figure is the company’s own, from a post I read in November 2026.
- StrongDMDescribed by Simon Willison, 7 February 2026
A three-person team with two rules: code must not be written by people, and must not be reviewed by people. Trust comes from end-to-end tests kept where the builders cannot see them. The team suggests spending at least $1,000 a day per engineer on AI usage. Read the post
- Stripe9 February 2026
More than a thousand changes merged each week contain no code written by a person. A person still reviews every one. Each agent works in a sealed developer machine, and most runs start from a chat message. Read the post
- OpenAI11 February 2026
An internal product of about a million lines, built in five months with no code written by hand, by a team that grew from three engineers to seven. The post describes the engineer’s job as designing the environment, stating intent and building feedback loops. Read the post
- Anthropic5 February 2026
Sixteen agents working side by side built a 100,000-line compiler over nearly 2,000 sessions, for just under $20,000. One of the author’s lessons: the tests have to be nearly perfect, or the agents solve the wrong problem. Read the post
- Anthropic24 March 2026
Agents mark their own work too kindly. The fix the post lands on is three agents: a planner, a builder, and a separate evaluator tuned to be skeptical. Read the post
- Ramp12 January 2026
Wrote up why it built its own background agent, called Inspect. It has since opened Inspect to teams outside Ramp in early access. Read the post
The download-button order followed through the story, and every number attached to it, is an invented example.