dispatching-parallel-agents

Dispatch every independent agent in a single message so they actually run in parallel. Each prompt needs four parts: the task as one goal, verbatim evidence like error text, the file boundary it may change, and a return contract describing what to hand back.

Use subagents when the user explicitly asks for fan-out and the work splits into two or more genuinely independent pieces. Fan-out buys wall-clock time and context headroom, but tokens scale linearly with agent count, so it never buys correctness.

Split by problem domain, not file count: each piece must be understandable without the others, and no two agents may write the same file. Related failures where one fix might resolve others, or exploratory work, are not real splits.

Not by default, since neither agent sees the other's diff. Use isolation: worktree to give each agent its own checkout when parallel writes would collide, accepting the setup time and disk cost.

Agent summaries are claims, not results; an agent may report tests passing without running them. Re-run the check yourself, run the full suite once at the end, and relay conclusions rather than transcripts.

There is no mid-course correction on a dispatched agent except SendMessage, and a bad prompt wastes the entire agent. A returned null means the agent died or was skipped and must be filtered before using results.