Think Tool

The think tool gives the agent a dedicated space for structured reasoning. Instead of immediately calling tools or generating code, the agent can pause to analyze information, plan multi-step approaches, and reason through complex problems.

This complements the Deep Agents write_todos tool (provided by TodoListMiddleware) — while write_todos tracks task progress externally, think enables internal deliberation before committing to actions.

When it’s used

The agent invokes think when it needs to:

• Analyze tool outputs — Process results from previous tool calls before deciding the next step
• Plan multi-step solutions — Break down a complex request into numbered action steps
• Evaluate trade-offs — Compare implementation approaches before committing
• Follow complex guidelines — Reason through project rules or coding standards
• Debug systematically — Formulate hypotheses about a bug before inspecting code

How it works

The think tool accepts a single thought parameter — a free-form string where the agent writes its reasoning. The tool doesn’t execute code, fetch data, or make changes. It simply creates space for deliberate analysis in the agent’s execution trace.

Agent: [think]
  The user wants to refactor the authentication module. Let me analyze:
  1. Current auth is in src/auth/ with 3 files: login.ts, session.ts, middleware.ts
  2. login.ts has a SQL injection vulnerability (string concatenation in query)
  3. session.ts uses in-memory store — won't work in multi-instance deployment
  4. middleware.ts is clean but could use rate limiting

  Plan:
  1. Fix SQL injection in login.ts first (security priority)
  2. Migrate session store to Redis adapter pattern
  3. Add rate limiting middleware
  4. Update tests for each change

Output

The tool returns the agent’s thought as a structured plan, which becomes part of the conversation context. This means:

• The agent can reference its own reasoning in subsequent turns
• The user can see the agent’s thought process in the conversation
• The plan guides the agent’s subsequent tool calls

Why it matters

Without think, agents tend to act impulsively — calling the first tool that seems relevant. The think tool encourages a “plan then act” workflow that produces more coherent multi-step solutions, especially for:

• Large refactoring tasks
• Debugging sessions with multiple hypotheses
• Architecture decisions with trade-offs
• Tasks that span multiple files