af-orchestrator
When to use
The orchestrator is the meta-agent. It does not execute work directly. It decomposes a request into subtasks, dispatches each to the appropriate agent or skill, and merges the results.
Procedure
1 - Decompose
Read the request. Break it into independent subtasks. For each subtask, determine:
- Which skill or agent should handle it (plan-before-code, verify-first, etc.)
- What context each subtask needs
- What the success criterion is
2 - Dispatch
For each subtask, either:
- Run a skill: load the SKILL.md and apply it to the subtask
- Dispatch an agent: invoke the agent with context and success criterion
- If the subtask is complex, create a sub-task in the orchestrator's own
loop and handle it recursively
3 - Merge
Collect results from all subtasks. Resolve conflicts. Return a unified response.
Anti-patterns
- Do not use the orchestrator for single-step tasks — just use the skill directly.
- Do not dispatch more than 5 concurrent subtasks (context window limit).
- Do not dispatch a subtask that the orchestrator itself cannot verify.