Ben ShapiroThe Factory Floor

The Factory Floor

Follow one order down a production line staffed by AI agents. Then turn the dial and watch where the people go.

A few software teams now let AI agents take a job from request to finished, tested change, with no person writing the code. The people on those teams work at the two ends of the line: deciding what gets built, and deciding what counts as done. This page follows one customer request down that line, then turns a dial through five levels so you can see who does what at each. It carries on from Robot in a Room.

Ben Shapiro · 14 November 2026 · About twelve minutes

Scroll to begin, or press →

The 3D floor could not start in this browser, so here is the full text on its own. Every step is below.

Opening

Pull back

Real nameSoftware factory

Last time: one robot in one room. Pull back, and that room is one workstation on a factory floor. A software factory is a production line where AI agents take a piece of work from request to finished, tested software. People run the line. They no longer do the work on it.

Stop 1 of 7

The order window

Real nameIntake, or triggers

Work used to start when a person sat down to do it. Here it starts when an order arrives: a ticket in your tracker, a message in chat, an alert at 3am, a job on a schedule. Nobody has to be at a keyboard.

Our order

A customer writes to support: “I can’t download my invoice.” Support logs it. That ticket is the order.

Stop 2 of 7

The blueprint desk

Real nameSpecification

Before anything is built, the order gets a blueprint: what is wanted, and how we will know it is right. On this line, the blueprint is the most valuable thing a person writes. A vague order produces confident, wrong work, quickly.

Our order

A product manager adds two lines. Put a “Download PDF” button on every invoice. The PDF must match the invoice on screen, on a phone and on a laptop.

For you

Your best people move from doing the work to describing it precisely.

Stop 3 of 7

The cell

Real nameSandbox: a sealed, disposable workspace

Each order gets its own sealed cell: a fresh copy of everything needed for the job, cut off from the live business. The robot can try, break and retry without touching anything real. When the job ends, the cell is wiped.

Our order

The cell holds a copy of the invoice system filled with made-up customers. Nothing in it can reach a real invoice.

Stop 4 of 7

The tool crib

Real nameShared tools and company context

Robots do not bring their own tools. They draw them from one shared crib: your systems, your documents, your house rules, each issued under permissions you set. Stock the crib once and every robot on the line can use it.

Our order

The robot draws three things: the invoice layout rules, the brand guide, and the PDF tool the company already uses.

Stop 5 of 7

Many cells

Real nameParallel agents and orchestration

One robot does one thing at a time. A factory runs many cells at once. A dispatcher hands out orders and makes sure no two cells take the same one. Big jobs are split up and worked in parallel, through the night.

Our order

While this cell builds the button, the next cell writes the help article for it. Every other cell is busy with another customer’s request.

Stop 6 of 7

The inspector and the test track

Real nameVerification

The builder never marks its own work. A separate inspector, who did not see it built, runs the result against tests the builder was never shown. Failures go straight back to the cell. This station sets the pace of the whole line, and how far you can trust it.

Our order

The inspector downloads 200 test invoices and compares each PDF with the screen. One with a very long address runs off the page. Back to the cell. On the second lap all 200 match.

For you

Spend here first. Every other station relies on this one to catch what is wrong.

Where the picture bends

If the inspector is the same kind of AI as the builder, the two can share blind spots. Who writes the tests matters as much as who runs them.

Stop 7 of 7

The gate and the control room

Real nameApproval policy and monitoring

Nothing leaves without passing the gate. People set its rules in advance: what ships on its own, what needs a signature, what never ships. Above the floor is the control room, where people watch throughput, quality and cost. When something stalls they fix the line, not the order.

Our order

A display change like this is on the “ships on its own” list, so it goes straight out. Had it changed what a customer is charged, it would have waited for a signature.

For you

Decide what needs a signature before the line runs.

Level 1 of 5

Hand tools

Real nameAutocomplete

A person does the work. AI finishes sentences and suggests the next line. The typing is faster. The factory is the same.

Our order

A developer builds the invoice button over two days, with AI finishing lines as they type.

For you

A company is not at one level. Each kind of work is. Internal tools might run at level 4 while anything touching customers’ money stays at 2.

Credit

The five-level ladder is Dan Shapiro’s idea, from January 2026. The factory names are mine.

Level 2 of 5

Power tools

In plain termsAn assistant, on request

AI produces whole pieces of work on request. A person reads and approves every one. The person is now the bottleneck.

Our order

AI writes the button in an hour. A developer reads every line before it goes out.

Level 3 of 5

The line

Real nameBackground agents

Orders run down the line unattended and arrive finished. People write the blueprints and sign at the gate. Most of their time now goes on review.

Our order

The line delivers the button finished. A developer approves it at the gate.

Level 4 of 5

The trusted line

In plain termsRoutine work ships by itself

This is the line you just toured. Routine orders ship without a signature, because the inspection has earned it. People sign only the risky ones and audit samples of the rest.

Our order

The button ships by itself. A person spot-checks one change in twenty like it.

Where the picture bends

The dial makes moving up look like one decision. Each notch is earned by building better inspection first, and the cost of running the line rises with it.

Level 5 of 5

Lights out

Real nameThe dark factory

No person writes or reads the work. People set the intent, design the tests and watch the scores. A three-person team at StrongDM has said in public that it works this way. It takes inspection you would stake the business on.

Our order

Nobody at the company watches it happen. The customer simply gets the button.

Beyond software

Other kinds of work

In plain termsA four-question test

Software went first because it has four things going for it. The requests are written down. A machine can check the result. The work can be rehearsed safely. Mistakes can be undone. Other work fits the factory as far as it shares those four.

For you

Run the four questions on a piece of your own work before you ask how to automate it.

Where the picture bends

The factory is working in software at a handful of teams, and early everywhere else. The three orders on screen are illustrations, not measurements.

The ending

What stays lit

In plain termsThe decisions people keep

At level 5 nobody works on the floor, but the building is still staffed. People decide what gets made, what counts as right, and what must never ship. Those decisions stay with people at every level.

For you

The dial is yours now. Where is your work today? What would have to be true to move it one notch?

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 floorReal nameWhat it does
01The order windowIntake, or triggersTurns a ticket, a message, an alert or a schedule into work.
02The blueprint deskSpecificationStates what is wanted and how it will be checked.
03The cellSandbox: a sealed, disposable workspaceGives each job a sealed, disposable copy of what it needs.
04The tool cribShared tools and company contextIssues your systems, documents and rules, under permissions you set.
05Many cellsParallel agents and orchestrationRuns many jobs at once and hands out the work.
06The inspector and the test trackVerificationChecks the result independently, against tests the builder never saw.
07The gate and the control roomApproval policy and monitoringDecides what ships, and shows people how the line is doing.

The dial, level by level

LevelNameWho does the workPeople on the floorWhy we trust it
1Hand toolsA person, with AI finishing their lines19A person did the work
2Power toolsAI, one piece at a time, when asked19A person read every line
3The lineAI, start to finish, unattended3A person approved it at the gate
4The trusted lineAI. Routine work ships by itself2The tests passed, and a person samples
5Lights outAI0The 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.
OrderWritten down?Machine can check?Safe to rehearse?Can be undone?Where it ends up
Match a supplier’s invoice to the purchase orderYesYes. The numbers agree or they do not.Yes, on a copy of the booksYes, until it is paidRuns through. People audit samples.
Answer a customer asking “where is my parcel?”YesPartly. Facts yes, tone no.Yes, as a draftNo, once it is sentStops at the gate. A person signs the unusual ones.
Decide next year’s pricesRarelyNoNoNot easilyLifted 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.