Knob

Autonomy level

You might say: levels of AI coding, delegation level, trust dial

How much of the work agents do with no person in the loop. Practitioners borrow the levels of driving automation, from a human who approves every character to a factory that turns specs into software with no human review. A software factory starts at level 3, with a person reading every diff, and aims at level 4. There, people write specs and check outcomes, while verification and the merge policy let low-risk work merge without a person reading the diff.

What it does: Sets the human's job at each step (writer, reviewer, spec writer) and so sets how strong the brakes and the verification must be.
Ask: At this level, who starts the run, who reads the diff, and who can merge?
Typical: The scale runs from manual level 0 to level 5, where no human reviews the code. Levels 3 and 4 separate human diff review from spec-and-outcome oversight.
Trouble: The team claims a higher level than its brakes support, and defects reach production unreviewed.

What to look for

Whether a human or an event starts each agent session
Whether every pull request waits for a human approval
Work that continues while the team is away

Catalogue: Level

Level 0: manual: The human writes the code. AI is at most a search engine or an occasional accepted suggestion.
Level 1: discrete tasks: The human hands small, separate jobs to the AI, such as a unit test or a docstring, and still writes the important code.
Level 2: pairing: The human pairs with an AI coding tool like a colleague, in a flow state, and still drives every session.
Level 3: human in the loop: Agents write the code in several sessions at once, and the human's day is reviewing their diffs. A software factory starts here, with human review on every pull request. Review of every diff makes it level 3, however many agents run.
Level 4: spec and schedule: The human writes specs and plans, sets schedules, and leaves the agents to work for hours, then checks the results. A mature factory runs here. Evidence carries low-risk work to merge, and people read the diffs only of high-risk changes. The human checks outcomes such as passing tests and scenarios instead of reading each step.
Level 5: dark factory: Specs go in and software comes out, with no human writing or reviewing the code. It is credible only where agents cannot game the evidence, such as holdout scenarios run against twins of outside services. It may cover only part of a codebase, and only a handful of small teams work this way. "Dark" means no human is needed on the floor, as in a robot-staffed factory.

Sources

Dan Shapiro, The Five Levels: from Spicy Autocomplete to the Dark Factory, IntroductionI’ve now seen dozens of companies struggling to put AI to work writing code, and each one has moved through five clear tiers of automation.

Use it on your project

teardown-apply: hold "Autonomy level" (autonomy-level) from the teardown at https://teardownatlas.com/software-factory/ against my project

Paste it into a coding agent that has the teardown skills.

An AI agent wrote this map with the teardown skill. It can contain errors: check the sources before you rely on a part. About and legal