dispatching-parallel-agentslisted

# Dispatching parallel agents Independent problems investigated one after another waste time and fill one context. Give each problem to a focused agent at the same time, then integrate the results yourself. Use the Agent tool. If delegation is unavailable (for example, you are already a subagent), handle the domains one after another, with a short summary after each. ## When it fits - Several failures with different root causes (different subsystems or test files). - Research or scouting across unrelated areas. - Multi-angle reviews: correctness, security, tests, performance. ## When it doesn't - The failures may share a cause. Investigate once first (systematic-debugging). - The tasks edit the same files or shared state. Parallel writers conflict. - You don't yet understand the problem well enough to split it. ## Process 1. **Group by independent domain.** Write down why each group can't affect the others. 2. **Write one focused prompt per agent**, self-contained, since it has none of your context: - **Scope:** exact files, tests, or the question. - **Goal:** what "done" means, with the verification command. - **Constraints:** read-only, or which files it may change; don't touch anything else. - **Output:** root cause, changes as a file list, commands run with results, open questions. 3. **Dispatch all agents in one message** (several Agent calls side by side), using the matching agent: `Explore` for recon, `reviewer` for review, `implementer` or `general