Orchestrate
From Kody, adapted for kody.exchange.
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)
- 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 checksor 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 Grok 4.5 /
composer-2.5-fastfor 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.
Every code-changing agent must push its branch and create or update the pull
request with Cursor Cloud ManagePullRequest (Kent C. Dodds account) before
finishing. @kentcdodds/cursor createAgent always enables those tools. Never
have Kody/workflows/GitHub create the initial PR as a substitute.