dispatching-parallel-agents

When you have multiple independent problems, dispatch one agent per problem domain and let them work concurrently. Each agent gets isolated context — never your session's history. This preserves your context for coordination. Use when 2+ independent tasks exist (different test files, subsystems, bugs). Use when each problem can be understood without context from others. Use when agents won't interfere (no shared state, no same-file edits). ```dot digraph when_to_use { "Multiple failures?" [shape=diamond]; "Independent?" [shape=diamond]; "Can parallelize?" [shape=diamond]; "Single agent" [shape=box]; "Parallel dispatch" [shape=box]; "Sequential agents" [shape=box];

  "Multiple failures?" -> "Independent?" [label="yes"];
  "Independent?" -> "Single agent" [label="no - related"];
  "Independent?" -> "Can parallelize?" [label="yes"];
  "Can parallelize?" -> "Parallel dispatch" [label="yes"];
  "Can parallelize?" -> "Sequential agents" [label="shared state"];
}
```

**Identify independent domains** — Group failures by what's broken. Each domain is one agent's scope. **Create focused prompts** — Each agent gets: specific scope (one file/subsystem), clear goal, constraints (don't change other code), expected output format. **Dispatch in parallel** — Use Task tool for all agents concurrently. **Review and integrate** — Read each summary, verify fixes don't conflict, run full test suite, integrate changes.

Prompt structure

Good agent prompts are focused, self-contained, and specific about output:

Fix the 3 failing tests in src/agents/agent-tool-abort.test.ts:

1. "should abort tool with partial output" - expects 'interrupted at'
2. "should handle mixed completed and aborted" - fast tool aborted
3. "should track pendingToolCount" - expects 3 results, gets 0

These are timing/race condition issues. Your task:
1. Read test file, understand what each verifies
2. Identify root cause
3. Fix (don't just increase timeouts)

Return: Summary of root cause and changes made.

Each agent gets narrow, self-contained scope. Never dispatch without specific error messages and test names. Always run full test suite after integrating all fixes. Spot-check agent work — agents can make systematic errors. "Fix all the tests" — too broad, agent gets lost. Be specific: one file or subsystem per agent. Dispatching agents for related failures — fix one might fix others. Investigate together first.