Which stage areyou really on?
Answer for what your team actually does today. Most teams rate themselves one stage too high, because they count things that are set up but not yet reliable.
Assisted
Your stage: 0, Assisted.
Delegated
- Every developer uses an agent that edits files and runs commands.
- A rules file exists at the repository root.
- One command runs all the tests.
- Agent-written PRs are labelled.
Harnessed
- Branch protection blocks red merges with no admin bypass, and has done so on a real PR.
- One validation command is shared by the agent and CI.
- At least one hook has blocked a write while you watched.
- A builder–checker loop runs with written stop rules.
- A tests-ship-with-features gate has failed a real PR.
Verified autonomy
- End-to-end tests run in CI, without a human, on the riskiest flows.
- Every merged agent PR has verification evidence or an UNVERIFIED marker.
- Golden evals run on a cadence with zero standing FAILs.
- Every lesson learned has a covering test.
- A lint ratchet has failed a real PR.
Managed agent teams
- The PR body is a generated evidence bundle.
- At least three scheduled routines run under a written policy, with known cost.
- PRs get opened while nobody is working.
- Human minutes per merged PR is measured and falling.
- Merge authority is spread across more than one person.
Self-improving system
- Every completed task ends with a reflect step that can open a mechanism PR.
- Eval FAILs open tasks automatically.
- Outcome metrics sit beside process metrics.
- A narrow autonomous-merge policy is in version control and audited.
- Every routine’s instructions live in the repository.
Start with the first check.
Answer for what your team does today. Your stage updates as you go.
Pick one agentic tool and give the whole team access.
Done when: every developer has run at least one session that edited a file and ran the tests.
Send your CEO or founder a link with your answers. When they finish the leadership edition, they’ll see exactly where you disagree. Your answers travel in the link itself; nothing is stored on a server.