dispatching-subagentslisted

# Dispatching Subagents ## Overview **Core principle:** Delegate to go faster and to keep your own context clean — but treat the host as a **budget**, not an infinite pool. The goal is *lower wall-clock*, not *maximum concurrency*. Ten agents that make the machine swap are slower than three that don't. Two wins from delegating, and they're different: - **Parallelism** — independent work runs at once, so wall-clock ≈ the slowest chain, not the sum. - **Context hygiene** — a subagent reads the haystack and hands you the needle; the file dumps never enter your context. Reach for this even when you *don't* need speed. ## Reach for a Subagent When - **Fan-out search / exploration** — "where is X handled", sweeping many files or naming conventions. Delegate; keep the conclusion, not the excerpts. - **Independent tasks** — spikes of two rival approaches, reviewing N files, building M targets — anything with no ordering dependency between the pieces. - **Adversarial verification** — spawn independent skeptics to try to *refute* a claim (behavior unchanged? bug real?). Diverse lenses catch what one pass misses. - **Anything that would flood your context** with output you won't keep. ## Don't When - **Steps depend on each other** — B needs A's output. Delegating then blocking on it buys nothing but overhead. - **It's one known lookup** — you know the file and symbol. Just read it. - **The work is smaller than the dispatch cost** — spawning an agent to save five seconds loses.