agent-team-runner

Files (2)

• SKILL.md
• validate_team_spec.py · scripts/

SKILL.md (first lines)

# Preconditions

- Read applicable `AGENTS.md` instructions.
- Locate the requested `team-spec.json`.
- Run the bundled validator before spawning any subagent.
- Refuse execution if validation errors remain.
- For strict validation warnings, either correct them or state why the risk is accepted.

# Execution protocol

1. Load the Team Spec and compute topological waves from `tasks[].depends_on`.
2. Respect `topology.max_threads`, `budget.max_agents`, and `topology.max_depth`.
3. Spawn only agents needed for the current wave.
4. Use the custom agent named by each task's agent profile when available; otherwise use the closest built-in `explorer`, `worker`, or `default` agent with the required sandbox.
5. Give each subagent a minimal task packet. Never delegate the full workflow to a child.
6. Run read-only tasks in parallel when independent.
7. Run write tasks concurrently only when their write scopes are disjoint and the Team Spec explicitly permits it.
8. Require each subagent to return `agent-result-v1` fields: status, summary, evidence, changes, validation, risks, assumptions, and recommended_next_action.
9. Validate outputs before unblocking dependent tasks.
10. Apply retries exactly as specified. Record every retry reason.
11. Execute gates after the listed tasks. A required failed gate blocks downstream work.
12. Require explicit human approval for a `human-approval` gate. Do not simulate approval.
13. Have the integration owner perform final integration and full validation when defined.
14. Write a run report under `agent-team/runs/<run-id>/run-report.json`.

# Failure behavior

- On ambiguous acceptance criteria: stop and escalate.
- On overlapping writes: stop affected tasks, preserve work, serialize or replan.
- On invalid agent output: request one schema-correction pass if retries allow.
- On deterministic command failure: do not repeat without a changed implementation or diagnosis.
- On permission denial: do not bypass the configured sandbox; escalate.
- On exhausted retries: mark the task failed and block its dependents.

# Final response

Report:

- run status
- completed, blocked, and failed tasks
- gates and results
- files changed
- validation commands and outcomes
- unresolved risks
- human decisions still required
- run report path

From EmbroiderSnow/Agent-Team-Design, MIT.

More from EmbroiderSnow/Agent-Team-Design