swarmlisted

# Swarm — Parallel Sub-Agent Fan-Out ## Why this exists Under load, agents drift back to serial execution — reading one file, then another, then another — even when the questions are independent. Serial reads waste wall-clock and the human's attention. This skill is the durable fix: when the pattern matches, fan out. ## When to invoke (all three true) 1. **2+ independent questions** — each answerable without another's result. 2. **Read-only or non-conflicting** — no shared file writes, no single-writer state (a git commit, a service restart, one config file two agents would both edit). 3. **Bounded synthesis** — you can hold the union of N short reports in context. **Trigger patterns:** multi-source research (web + repo + package registry + docs) → one agent per source; multi-target verification (logs + config + git + db state) → one agent per target; 3+ independent file reads → one agent per file (or grouped); a verifier cluster after a fix → fan out across measurement angles. ## When NOT to Sequential dependencies (each step needs the prior result); shared writes (race conditions); single-writer operations (restart, commit); trivial single-tool work (overhead exceeds benefit). ## How Use the Agent tool with several calls **in one message** so they run in parallel: ``` Agent({ description: "...", subagent_type: "general-purpose", prompt: "<self-contained brief>" }) Agent({ description: "...", subagent_type: "general-purpose", prompt: "<self-contained brief>" })