dispatching-parallel-agents

Dispatching Parallel Agents

When to Parallelize

Tasks are safe to parallelize when:

• They touch different files (no merge conflicts)
• They have no shared state (no race conditions)
• They are independently verifiable (each has its own test)
• The order doesn't matter (no dependency chain)

Dispatch Pattern

1. Identify Independent Tasks

From the plan, group tasks by independence. Tasks that modify the same files are NOT independent.

2. Craft Parallel Prompts

Each agent gets:

• Its specific task with full context
• Architecture constraints relevant to its scope
• Clear file boundaries (what it owns, what it must not touch)
• Its own verification command

3. Launch Simultaneously

Use multiple Agent tool calls in a single message. This is critical — sequential launches waste time.

4. Collect and Review

When all agents complete:

• Review each result independently
• Check for unintended overlaps
• Run the full test suite (not just individual tests)
• Verify the combined changes work together

Anti-Patterns

• Parallelizing tasks that share files (creates merge conflicts)
• Launching agents sequentially when they could be parallel
• Not running integration tests after combining parallel results
• Giving agents overlapping file ownership

Example

Two bugs in separate modules:

Agent 1: Fix user service validation bug (src/services/user.ts, tests/user.test.ts)
Agent 2: Fix order API pagination (src/routes/orders.ts, tests/orders.test.ts)

These are safe to parallelize — different files, different concerns.