---
name: use-multi-agent
description: Coordinate multiple agents for parallel, non-blocking work. Use when the user explicitly asks for subagents, delegation, or parallel agent work and the task contains independent research, exploration, verification, or isolated sidecar execution that can run in parallel while blocking work stays local.
---

# Use Multi Agent

Plan locally first, then delegate only the work that is safe to parallelize.
Keep the critical path moving in the main thread.

## Required operating rules

- Identify the immediate blocking task before spawning anything.
- Keep blocking or tightly coupled work in the main thread.
- Spawn subagents only for independent, bounded, materially useful tasks.
- Prefer delegation for research, codebase exploration, inventory gathering, comparison, verification, and isolated implementation slices with a disjoint write scope.
- Do not delegate speculative work or tasks that the main thread needs immediately for the very next action.
- While subagents run, do non-overlapping local work.
- Wait only when blocked on a delegated result.
- Integrate returned results quickly and close agents that are no longer needed.

## Delegation checklist

1. Write a short local plan first.
2. Mark which step is on the critical path.
3. Separate sidecar tasks that can run independently.
4. Give each subagent a concrete deliverable.
5. For coding tasks, assign file or module ownership explicitly.
6. Tell coding subagents they are not alone in the codebase and must not revert others' changes.
7. Avoid duplicate delegation.
8. Summarize what was delegated and why when reporting progress.

## Good delegation targets

- Repo structure or dependency discovery
- Reading docs or references
- Comparing alternative approaches
- Verifying a suspected bug or risk
- Implementing a clearly isolated code change
- Running independent validation while the main thread keeps implementing

## Poor delegation targets

- The next blocking decision the main thread needs right now
- Broad, ambiguous "go solve this whole thing" requests
- Overlapping edits in the same files without clear ownership
- Re-running work already being handled locally

## Output expectation

When you use subagents, keep the user's progress readable:

- say what stayed local
- say what was delegated
- say what came back
- say how the results changed the next action