Adoption levels
You don’t install this system all at once. Start at the level that matches your trust and your time, and climb. Each level is useful on its own; each builds on the one before.
Level 0 — Read the model, change nothing
Just internalize the architecture: the four roles, the bands, the security spine. Even with zero automation, the mental model improves how you triage by hand. Cost: one read. Value: a clearer operating model.
Level 1 — Manual playbooks (Band B, human drives)
Run the setup interview and use the lifecycle playbooks as checklists in supervised sessions. The agent helps review a PR, triage an issue, run a release — but you’re in the loop for everything. Cost: setup + your attention per session. Value: rigor and consistency you didn’t have before. This is where most projects should start.
Level 2 — The scoreboard + digests (Band A reads, Band B acts)
Add the read-only autonomous layer: a triage scoreboard that ranks open work, a “what’s waiting on us” digest, a chat Watcher that captures reports. Nothing acts on the world yet — these just surface what needs attention so your Band-B sessions are sharper. Cost: a few scheduled jobs. Value: nothing falls through the cracks.
Level 3 — Mechanical autonomous actions (Band A, watchdogged)
Promote the safe, mechanical actions to autonomous, each with a watchdog: deterministic label-sync and the strict shipped-issue auto-close. Follow the autonomy ladder — small batch, verify, expand, watch. Cost: building the deterministic scripts + the watchdog. Value: the routine maintenance runs itself, verifiably.
Level 4 — Full pipeline
Deep PR review and builds run as Band-B sessions feeding a release pipeline whose merges stay Band B; the contributor trust ledger weights review; community monitoring and contributor recognition run continuously; a fleet heartbeat watches the whole thing. The maintainer’s job becomes decisions and taste, not mechanics. Cost: the full build. Value: an agent co-maintainer.
The climb is the point
Each level earns the next. Don’t jump to Level 3 autonomy for an action you haven’t run manually at Level 1 — you won’t have the evidence base or the watchdog to make it safe. The system is designed to be grown into, and to stay safe at every height.
| Level | What’s autonomous | What’s human |
|---|---|---|
| 0 | nothing | everything |
| 1 | nothing | everything (agent-assisted) |
| 2 | reads, ranking, capture, digests | all actions |
| 3 | + mechanical reversible actions (watchdogged) | judgment, merges, public voice |
| 4 | + the routine pipeline | decisions, taste, the irreversible calls |
Which skills fit which shape
Shape comes from the first questions of the setup interview. The table lists where each shape usually starts. A person confirms every block marked as not applicable.
| Shape | Skills | Blocks usually marked |
|---|---|---|
| public with outside reporters and tags | action-watchdog, coding-principles, contributor-trust, issue-autoclose, issue-capture, issue-close-proposal, issue-triage, label-sync, pr-deep-review, pr-triage, project-system-setup, release-announce, release-pipeline, triage-scoreboard |
none |
| private team with a separate tracker | project-system-setup, triage-scoreboard, issue-triage |
community, issue_capture, labels, trust |
| private solo | project-system-setup, coding-principles, pr-deep-review, release-pipeline |
community, issue_capture, trust |
The first job for a private team is triage-scoreboard, plus issue-triage run judgment-only by a
person, at Level 2. That run writes judgments and skips replies, labels and sprint-candidate marks.
The vulnerability divert still runs. issue-triage does no backfill, so that run reaches the
existing backlog only for the issues a person names.
A team with pull requests adds coding-principles, pr-triage, pr-deep-review and
release-pipeline once Level 2 is steady.
Related: autonomy ladder · setup interview.