Agent Handbook (Development Autonomy)¶
Date: 2026-02-08 Status: Draft
This file defines practical instructions that agents can follow during autonomous development.
Scope¶
- Safe autonomous development in this repository.
- No production-impacting actions without explicit approval.
Environments and Deployment Modes¶
- Development environment is default for iterative work.
- Production changes require explicit human approval and change context.
- Keep deploy command references here (or link to runbook) with clear
devvsprodexamples.
Logs and Diagnostics¶
- Prefer service-level logs first, then infrastructure logs.
- Standard workflow:
- Identify service and environment.
- Pull recent logs with bounded time windows.
- Correlate by request ID/run ID/tenant/workspace.
- Record concrete findings in the work summary.
Git Collaboration Workflow¶
- Branch naming:
codex/<scope>-<short-description>. - Keep commits small and reviewable.
- Do not amend shared commits unless requested.
- Do not force-push shared branches unless requested.
- If unexpected repo changes are detected, stop and ask before proceeding.
Multi-Agent Structure (Recommended)¶
- Use layered instructions:
- Root-level
AGENTS.mdfor global repository policy. - Service-level
services/<service>/AGENTS.mdfor local commands, test workflows, and caveats. - Service-level instructions should not conflict with root safety rules.
- Each service-local file should link back to shared contracts and specs in
/docs.
Change Safety Rules¶
- Prefer additive changes over broad refactors.
- Keep API and contract compatibility unless a linked ADR approves breaking change.
- Update relevant docs with any behavior change.
- Use PR checklist gates for docs impact and traceability to reduce drift over time.
Definition of Complete (for agent tasks)¶
- Code change implemented.
- Tests updated and run for affected scope.
- Docs/specs/ADRs updated as needed.
- Rollout risks and validation steps documented.