<tool name="agent_task">
**When to use the agent_task tool:**
- Open-ended searches requiring multiple rounds of exploration
- Complex multi-step research tasks where you're uncertain about the approach
- When searching for a keyword or file and not confident you'll find the right match in first few tries
- Tasks that would consume excessive context if done in the main conversation, such as reading websites when looking for a small piece of information.
- Specialized tasks that match a specific agent type's expertise

**When NOT to use the `agent_task` tool:**
- You know specific file paths → use `read_file_lines` directly
- Searching for a specific class definition → use `grep_files`
- Searching code within a specific file or set of 2-3 files → use `grep_files` and `read_file_lines`
- Simple, focused tasks → do them inline yourself

**How to use the `agent_task` tool:**
- Agents run autonomously and return results in one message
- You cannot interact with them after launching - they are stateless
- Clearly specify whether the agent should research or write code
- Provide detailed, comprehensive instructions in the prompt parameter
- You can launch multiple agents in parallel for independent tasks
- Agent results should generally be trusted

**Available agent types:**
- `introspector`: Specialized agent for exploring elisp APIs and the state of the Emacs instance in which you are running.  Has access to various elisp introspection tools
- `researcher`: Focused on searching, reading and analyzing without modifications (Tools: `read_file_lines`, `glob_files`, `grep_files`, `search_web`, `read_url`)
</tool>