Multi-Agent

Decompose complex tasks into parallel subagent work. Use when multiple independent concerns need to be addressed simultaneously.

Recommended model: Opus (orchestration reasoning)

When to Use

Use multi-agent dispatch when:

• 3+ independent files or modules need changes with no shared state
• Multiple concerns need investigation at the same time (security + performance + correctness)
• A task has clear subtask boundaries that can be verified independently

Do NOT dispatch agents when:

• Subtasks share files they will both write to (race condition on changes)
• One subtask's output is required as input for another (sequence instead)
• The task is simple enough for a single focused response

Rules

1. Identify independent subtasks

Before dispatching, list subtasks and verify independence:

• No shared file writes
• No data dependency between subtasks
• Each subtask has a clear, verifiable output

2. Assign model per subtask

3. Present dispatch plan and wait for approval

Show the plan before dispatching:

Parallel dispatch plan:
  Agent 1 [Sonnet]: Implement user authentication middleware — output: auth.ts
  Agent 2 [Sonnet]: Write integration tests for /v1/users — output: users.test.ts
  Agent 3 [Haiku]:  Review API design conventions in routes/v1/ — output: findings list

Proceed? [y / n / modify]

Wait for explicit y before dispatching.

4. After agents complete

• Show each agent's output summary
• Check for conflicting changes (same file edited by multiple agents)
• Run integration verification: tests, type check, lint
• If conflicts exist: resolve manually before declaring done

5. Never dispatch subtasks with shared write targets

If two agents need to edit the same file, sequence them — don't parallelize.

Dispatch Format

When dispatching, give each agent:

• Clear scope (which files, which concern)
• Expected output format
• Constraints (follow ts-api / go-api standards, no scope creep)
• Model to use

What NOT to do

• Do not dispatch more than 5 agents in one batch
• Do not dispatch without user approval
• Do not assume subtasks are independent without checking for shared state
• Do not merge agent outputs without verifying no conflicts