Agent Tools

Agent

Spawns a nested agent to complete a focused sub-task. The sub-agent gets a fresh conversation history and runs the full agent loop independently.

{
  "tool": "Agent",
  "input": {
    "task": "Read all TypeScript files in src/components/ and write a summary of each component's props interface to components.md"
  }
}

The sub-agent has access to all the same tools as the parent. Its final response is returned as the tool result.

Nesting limit: 3 levels deep.

Use sub-agents to:

• Parallelize independent research tasks
• Isolate a focused sub-task from the main conversation context
• Prevent tool output from a large scan from filling the parent’s context window

---

Task tools

Task tools let the agent create and track a structured to-do list, visible to the user in the terminal.

TaskCreate

{
  "tool": "TaskCreate",
  "input": {
    "subject": "Add error handling to the login flow",
    "description": "Wrap the auth call in try/catch and show a user-friendly error message"
  }
}

TaskList

{"tool": "TaskList", "input": {}}

Returns all tasks with their status (pending, in_progress, completed).

TaskGet

{"tool": "TaskGet", "input": {"taskId": "1"}}

Returns full details of a single task including description and dependencies.

TaskUpdate

{
  "tool": "TaskUpdate",
  "input": {
    "taskId": "1",
    "status": "in_progress"
  }
}

Valid statuses: pending, in_progress, completed, deleted.

Can also update subject, description, set dependencies with addBlockedBy/addBlocks.

TaskStop / TaskOutput

Used by sub-agents to stop a background task or retrieve its output. Generally not called directly by the LLM.