What a dark factory is, and why the line is not one.
On the factory floor it is the plant that runs with the lights out, because nobody is there to see anything. In software, it is the code nobody reads before it goes into production.
Published
Specs in, software out.
The name already existed in industry, at the Fanuc plant where robots build robots. Dan Shapiro gave that name to the top rung of coding autonomy, and the definition is literal.
A dark factory is not another word for a coding agent, and not another word for automated review. A dark factory is defined by the absence of human reading, not the presence of AI.
There are six rungs, from autocomplete to black box.
Shapiro borrowed the scale from the levels of driving automation the NHTSA published in 2013. Five levels, six rungs: the count starts at zero, as it does in the scale it copies. Each rung takes one more piece of the work out of human hands.
- 0Writing
- Not a character is written to disk without human approval. AI comes in as a faster search, at most.
- 1Delegating
- Discrete tasks go to the model: a unit test, a docstring. The work still moves at the speed of whoever is typing.
- 2Pairing
- The model comes in as the pair and takes all the boring work. Shapiro puts 90% of self-described AI-native developers here.
- 3Reviewing
- The model writes almost everything and the whole day turns into diff review. This is the rung almost everyone tops out on.
- 4Specifying
- What is left is writing and arguing over the spec, crafting the rules and checking the result. Shapiro places himself here.
- 5Nobody reads
- A black box that turns specs into software. Shapiro says he knows a handful of people doing this, in teams of under five.
The line runs at level 4, and that is where it was built to stop.
The line is operated by AI and supervised by humans. Moving to level 5 is giving up the reading.
- Who understands
- The knowledge base is a station on the line. Codebase, documentation and anonymized data go into it before a single new line is written.
- Who finds it
- Observability is another station. When something breaks at three in the morning, the damage depends on how long it takes to find the cause, and the system ships instrumented end to end.
- Who answers
- Every action is logged and attributable. Months later, the record still shows who decided what.
- Who decides
- Your team sets the limit at each station, and keeps changing it after the deployment.
None of this is a prediction about level 5. It is the position the line operates from today.
Four months of code nobody read.
In July 2025, Dex Horthy set the model to write the code and kept the team out of review. Four months later they shut it down and threw the whole system out.
Production went down, and however hard the team pushed, the model could not find the root cause. It took days of digging through the code to surface a primary key routed wrongly through the whole system, and three more weeks to re-onboard engineers into a codebase nobody had ever read.
Four months of writing fitted inside an automatic loop; sustaining what had been written did not. Sustainment is a station on the line for that reason.
The lights stay on.
The line takes work off your team that does not need a person. The line does not take the people out of the room. The code stays legible to whoever maintains it later.
How legacy is modernized without stopping the business
- Dan Shapiro, “The Five Levels: from Spicy Autocomplete to the Dark Factory”, 23 January 2026The six rungs, the name, and the comparison with the levels of driving automation.
- Gergely Orosz, “Context engineering with Dex Horthy”, The Pragmatic Engineer, 15 July 2026The four months, the shutdown, and the three weeks of re-onboarding.