Agent Fan-out

Triggers automatically when relevant; invoke directly with /agent-fanout or ask, e.g. "audit all modules in parallel".

Context is your scarcest resource; quality degrades as it fills with noise. Delegation is a context firewall: the subagent burns ITS context reading forty files and returns you a ten-line conclusion.

Decision tree

Does answering require reading more than you need to retain?
├─ NO (one fact, known location) → do it yourself, directly
└─ YES
   ├─ one independent investigation → one agent
   ├─ several INDEPENDENT ones      → several agents IN PARALLEL,
   │                                  launched in one message
   └─ unknown-size discovery        → waves until 2 consecutive waves
                                      find nothing new (loop-until-dry)

Delegation contract — every agent prompt contains three things

1. TASK with boundaries — "examine X, ignore Y". Unbounded agents wander.
2. RETURN SHAPE — structured data, not prose: "return JSON {files, risk, evidence}". The agent's final text is a return value, not a message to a human.
3. DONE CRITERION — when to stop. Without it agents stop too early or never.

Flow rules

• Scout inline first (cheap listing), then fan out over the discovered work-list. You need the list's shape, not the whole task's shape.
• Pipeline by default: items flow through stages independently. Synchronize all agents only when a stage truly needs ALL prior results — dedup across the full set, early-exit on zero findings, cross-comparison.
• Loop-until-dry convergence: dedupe new findings against everything SEEN, not just accepted ones — otherwise rejected findings reappear every wave and the loop never terminates.
• Scale to the ask: "check" → 2–3 agents; "thorough audit" → a large fleet plus an adversarial-verify layer.

Discipline after delegating

• Delegated a search → never redo it yourself. Wait.
• Relay agent conclusions to the user — reports come to you, not them.
• Keep conclusions, discard raw material; that's what the firewall is for.
• Never cap silently: if you sampled or limited coverage, say what was dropped.

Relationship to other skills

• recon-first produces the work-list this skill fans out over.
• adversarial-verify filters what the fleet finds before anyone acts on it.