Why this playbook
Most agent demos look magical until they touch your PR queue. This short playbook is the version that survived friction with real teams.
Start with latency, not agents
If CI feels slow, nothing else will stick. Make feedback loops snappy first:
- Cache the boring parts (deps, docker layers, test artifacts)
- Shrink PR diff size with guardrails
- Add flake quarantine instead of toggling tests
Four agents that pull their weight
- Product Owner Agent: tight backlog items with acceptance criteria and contradictions highlighted.
- Architect Agent: paved-path patterns and trade-offs with pointers to prior decisions.
- Developer Agent: repo-aware scaffolds + tests sized for small, reviewable diffs.
- Testing Agent: skeleton tests and mocks that shift defect discovery to pre-merge.
Friendly defaults that avoid backlash
- Suggest patches on a side branch; let humans opt-in.
- Cap max diff size per agent run.
- Keep production changes behind a feature flag with a human approver.
Rollout cadence that works
Start with one squad. Publish the playbook as you go. Measure time-in-queue per lane and PR size distribution so the story you tell is obvious.
If you want the longer version (with templates and repo bootstrap), ping me and I’ll send the starter kit.