Swarm

Swarm (/swarm)

When to use

Apply this skill whenever the user sends /swarm (with or without extra text in the same message). Treat it as an explicit command to parallelize the request across multiple subagents.

Rules

1. Minimum agents: n ≥ 2. Never satisfy /swarm with a single subagent or a single serial plan without delegation.
2. Scale n with complexity — bias upward; when in doubt, add another agent rather than overload one.
   • Small / localized (one file, one bug, narrow question): n = 2–3 (e.g. implement + verify, or explore + implement).
   • Medium (feature touching several files, API + UI, tests): n = 3–5 (e.g. backend, frontend, tests, docs or integration).
   • Large (refactor, multi-package, audit, unfamiliar codebase): n = 5–8+ (partition by subsystem, directory, concern, or risk).
   • Very large / open-ended: split by independent workstreams so agents do not block each other; n as high as is useful within practical limits (each Task must have a crisp prompt and clear deliverable).
3. How to run the swarm: Use the host agent's native subagent or delegation tool to launch separate workers in parallel (same turn if possible). If the host has no delegation capability, explain that /swarm is unsupported instead of simulating parallelism. Each worker gets:
   • A narrow scope and concrete output (files to touch, questions to answer, or artifacts to produce).
   • No duplicate ownership of the same file unless one agent reads and another writes with an explicit handoff in your merge step.
4. You (orchestrator) must:
   • Briefly state n and why that count matches complexity (one short paragraph).
   • After subagents return, merge: resolve conflicts, unify design, run or request a single validation pass, and produce one coherent answer for the user.
5. If the user also gave a task description in the /swarm message, that description is the goal; decomposition is yours.

Anti-patterns

• One giant Task prompt that does everything.
• n = 1 or purely sequential subagents when work could run in parallel.
• Skipping synthesis so the user sees disconnected partial reports.

Model hint

Use a general-purpose worker unless the host offers a more appropriate specialist. Use isolated worktrees for concurrent code edits when the host supports them.