Launch multiple subagents to execute independent tasks concurrently.

USING PARALLEL TASKS YIELDS 2-5x EFFICIENCY GAINS FOR INDEPENDENT WORK.

## When to Use

- Run 2-10 independent research/coding tasks at once
- Explore multiple code areas simultaneously
- Execute independent file operations in parallel
- Launch multiple reviewers with different focus areas
- Speed up multi-step workflows where steps are independent

## When NOT to Use

- Tasks that depend on each other's output
- Sequential workflows where order matters
- Single tasks (use `task` tool instead)
- Tasks requiring back-and-forth with user

## Subagent Types

| Type | Best For | Model |
|------|----------|-------|
| `coder` | Multi-file edits, features | Claude Sonnet |
| `researcher` | Code analysis (read-only) | Claude Sonnet |
| `reviewer` | Bug detection, security | Claude Sonnet |
| `planner` | Implementation plans | Claude Sonnet |
| `finder` | Fast code search | Gemini Flash |
| `oracle` | Deep reasoning | o1 |
| `librarian` | Multi-repo research | Claude Sonnet |
| `auto` | Router picks best | Varies |

## Parameters

- `tasks`: Array of 1-10 tasks, each with:
  - `subagent_type`: Agent type or "auto" for routing
  - `prompt`: Detailed task instructions
  - `description`: Short 3-5 word summary
- `fail_fast`: (optional) Cancel remaining if one fails

## Example

```json
{
  "tasks": [
    {
      "subagent_type": "researcher",
      "prompt": "Find all API endpoints in src/api/ and document their parameters",
      "description": "Find API endpoints"
    },
    {
      "subagent_type": "researcher",
      "prompt": "Analyze the authentication flow in src/auth/",
      "description": "Analyze auth flow"
    },
    {
      "subagent_type": "finder",
      "prompt": "Find all test files and list their test functions",
      "description": "Find test files"
    }
  ],
  "fail_fast": false
}
```

## Tips

1. **Be specific** - Each subagent has no shared context
2. **Include file hints** - Speed up search with path hints
3. **Use appropriate types** - `finder` for speed, `oracle` for depth
4. **Group related work** - Tasks with shared context run better together
5. **10 task limit** - Additional tasks are ignored
