| name | Implement (multi-agent loop) |
| description | Orchestrate an implement -> review -> fix loop with subagents until reviewers sign off. |
| triggers | ["implement","build feature","multi-agent","delegate","implement loop"] |
Implement (multi-agent loop)
Use this when you (the orchestrator) should implement a non-trivial change by fanning
work out to subagents and iterating until the change is clean. You drive the loop with
the delegate tool; subagents work in isolated copies and their changes await the
user's merge review.
Loop
- Plan & split. Break the task into independent, self-contained chunks. Note the
acceptance criteria you will check against.
- Implement.
delegate one implementer per chunk (parallel — pass them in a
single delegate call). Give each a precise, standalone instruction.
- Review. When the implementer summaries return,
delegate a reviewer (one, or
one per area) to review the changes for correctness, security and quality. Tell the
reviewer exactly which files/areas to look at.
- Fix. Collect the reviewer findings. If there are any blocking findings,
delegate an implementer to fix them (reference the specific findings).
- Repeat steps 3–4 until reviewers report no blocking findings, or you reach the
round limit. Default to 3 rounds; if the user asked for an "effort N" (1–5),
use N rounds.
- Summarize. Report what changed, the final review verdict, and that the changes
are staged for the user's merge review.
Notes
- Keep your own context clean: integrate the returned summaries, not transcripts.
- Do not exceed the team limits (parallelism / subagent count / turn budget) — they are
enforced; design your fan-out to fit.
- If a chunk turns out to depend on another, sequence those
delegate calls instead of
running them in parallel.