Multi-Agent Orchestration

When to use

• Execute authorized agent work with explicit ownership and integration checks.

Not for: breakdown for decomposition; implement for single-agent work.

Steps

1. Confirm delegation is authorized and supported; use it only when isolation or concurrency earns the overhead.

2. Separate dependencies and independent tasks; assign file and state ownership so writers do not collide.

3. Provide goals, inputs, decisions, boundaries and verification targets; describe actual shared context and filesystem capabilities.

4. Inspect returned diffs and rerun checks before integration; reconcile assumptions and verify the whole result.

Verify

• Ownership and dependencies are explicit; integration is checked independently of agent claims.

References

• engineering-principles.md — shared decision boundaries, scope and verification.
• references/dispatch-patterns.md — when to fan out vs sequence, context-construction templates, review-checkpoint patterns, common failure modes (race conditions, context-starved subagents, unverified seams).