Your autonomy level is a property of the repo.
Not of you, and not of the model. Same engineer, same week, drop into a legacy codebase with no types and a flaky test suite and you are back to pairing instantly. Only the repository changed.
Twenty questions about what your team actually does while working with coding agents. At the end you get the level your repository can safely support, and the single thing holding you there.
Answer for what is true today. Not what is on the roadmap, and not what one person does on their own branch.
Your answers stay in this browser. Nothing is sent anywhere and there is no account to create.
What it measures
Five areas, four questions each. The lowest one sets your level, because a strong spec practice does not compensate for a verification stack that does not exist.
Spec
What "done" means before you prompt.
Context
What the repo teaches every session.
Harness
What the agent can run, see, and is allowed to touch.
Verification
What tells the agent it is wrong, without you.
Compounding
Whether lessons become gates.
Where it comes from
The ladder is Dan Shapiro's. The pillars are Alfonso Graziano's. Gating the ladder on verification, and treating verification as a pillar rather than a review step, is mine.
Before you start
- What are the five levels of AI coding autonomy?
- Dan Shapiro adapted the self-driving levels to AI-assisted development. Level 0 is manual, level 1 is discrete delegated tasks, level 2 is pairing with the model in a flow loop, level 3 is supervising several agent branches and reviewing diffs, level 4 is writing specs and leaving, and level 5 is a dark factory where specs go in and software comes out.
- Why does this measure my repository rather than my team?
- Because each transition up the ladder is gated by one question: how cheaply and reliably can your environment tell the agent it is wrong. That is verification machinery, and it is something you build. The same engineer with the same model operates at level 4 on a repo with strict types and behavioural CI, and drops to level 2 in a legacy codebase with a flaky test suite. Only the repository changed.
- Why is my level the lowest area rather than the average?
- Because gating means gating. A strong spec practice does not compensate for an absent verification stack, it just produces well specified code that nobody can trust. Averaging the five areas would hide exactly the failure the ladder is about, so the lowest area sets the level.
- How long does it take?
- Twenty questions, about five to six minutes. Answers stay in your browser, nothing is uploaded, and there is no account to create.