| name | agent-orchestration |
| description | Use when a plan contains independent research, implementation or review tasks that benefit from fresh-context subagents or parallel execution. Do not parallelise tasks that modify the same state without a merge protocol. |
Agent Orchestration
Eligibility
Parallel execution requires:
- at least two independent tasks;
- explicit input and output contracts;
- separate files or non-conflicting state;
- bounded context for each agent;
- named integration owner;
- deterministic verification.
Procedure
- Choose the smallest number of agents.
- Give each agent only the context needed for its task.
- Require an artifact, evidence and unresolved issues.
- Use two-stage review for implementation:
- specification compliance;
- engineering quality.
- Merge sequentially through the integration owner.
- Re-run whole-system verification after merge.
Prefer installed Superpowers subagent-driven-development or
dispatching-parallel-agents when compatible.