Think Tool
think gives the agent a dedicated space to reason before it acts. It doesn’t execute code, fetch data, or change state — it just writes the agent’s reasoning into the conversation trace so it (and you) can see the plan.
Complements deepagents’ write_todos (external task tracking) — think is internal deliberation before committing.
When the agent uses it
Section titled “When the agent uses it”- Analyze tool outputs before deciding the next step
- Break a complex request into numbered actions
- Compare implementation approaches
- Reason through project rules or coding standards
- Formulate hypotheses before inspecting code
Without think, agents tend to act impulsively — call the first tool that seems relevant. think encourages “plan → act” and produces more coherent multi-step solutions.
Single argument, thought: string. Free-form. Example agent output:
[think]The user wants to refactor auth. Analysis: 1. login.ts: SQL injection via string concatenation. 2. session.ts: in-memory store — breaks in multi-instance deployment. 3. middleware.ts: clean; could add rate limiting.
Plan: 1. Fix SQL injection first (security priority). 2. Migrate session store to Redis-adapter pattern. 3. Add rate limiting. 4. Update tests per change.The plan becomes part of the conversation context — the agent can reference it in subsequent turns, and you see it in the chat panel.