| name | orchestrate |
| description | Orchestrate sub-agents for large tasks inside a single environment: plan, delegate coding to cheap/fast models, parallelize the critical path, keep reviews lean, and close every loop. Use when acting as an orchestrator or writing a kickoff for one. Prefer implement when fan-out does not clearly pay. For multi-environment fleets, use conduct instead.
|
Orchestrate
Fan out sub-agents inside one environment (shared checkout). For separate
environments/PRs, use conduct.
Two modes: be the orchestrator, or spawn one (smarter model) if you are
optimized for cheap/fast execution.
Defaults (override only if the user says so)
- Model. Use the latest Grok model unless the user requests another.
- Prefer implement. Fan out only when non-conflicting workstreams clearly
beat one implementer. Sequential slices → implement (or one implementer).
- No orchestration theater. Ban review → recheck → final → CI-watch chains
per slice. One hard-to-reverse review before merge.
- One CI wait path — parent
gh pr checks or a ci-watcher, not both.
- Targeted tests while iterating; full validate before ready-for-review.
- Close loops yourself — don't leave humans as the relay.
Role
- Plan, delegate, integrate, ship. Bulk-code only when fan-out costs more.
- Frontier model orchestrates; cheap/fast models implement (prefer the latest
Grok model for mechanical work).
- Critical path first; parallelize non-conflicting files; serialize shared
ones.
- You do final QA. Never declare done from sub-agent claims.
Fan-out (when it pays)
- One implementer per independent vertical slice
- Cheap sweeps (parity / perf / errors)
- Audit → prioritized cleanup, then parallel cleanup agents
- One independent "hard-to-reverse / security / perf holes" review before merge
Kickoff (when spawning an orchestrator)
Keep it short: goals + constraints + out-of-scope; "you orchestrate, don't bulk
code"; preferred implementer model; single-environment (not conduct); done =
falsifiable. Point at this skill.
