flpbalada Agent Harness Orchestration
Coordinate many coding agents without corrupting shared state.
Humans steer. Agents execute. Harness gives context, isolation, and feedback.
Goal
Turn tasks into isolated, verifiable, review-ready work.
Rules
- One task per branch or worktree.
- One owner per path when possible.
- Do not let agents edit the same files at the same time.
- Sequence migrations before dependent feature work.
- Keep shared refactors separate from feature work.
- Require evidence, not claims.
- Capture follow-ups in repo docs, tests, or tracker.
Task Packet
Give each agent:
- Objective.
- Scope and owned paths.
- Out of scope.
- Acceptance criteria.
- Verification commands.
- Required context:
AGENTS.md, docs, specs, issues. - Branch or worktree name.
- Output contract.
Flow
- Intake task.
- Split work and mark dependencies.
- Start only unblocked tasks.
- Create isolated workspace.
- Assign packet.
- Agent edits and verifies.
- Supervisor reviews evidence.
- Rework, prepare PR, or capture follow-ups.
Verification Loop
- Reproduce bug or baseline when relevant.
- Make smallest useful change.
- Run targeted checks.
- Run broader checks.
- Smoke test user path or API.
- Report exact commands and results.
Escalate missing secrets, unsafe migrations, flaky infra, or unclear product judgment.
Failure Modes
- Shared mutable workspace.
- Vague task packet.
- No mechanical verification.
- Hidden architecture drift.
- Lost follow-up work.
Output
## Agent Work Summary
- Ready: [tasks]
- Blocked: [tasks and reason]
- Conflicts: [paths or none]
- Verification: [commands and results]
- Follow-ups: [docs, tests, tasks]