---

name: gemini-agentic-react description: strategies for implementing the ReAct (Reason and Act) paradigm. Use this to enable Gemini to use external tools, APIs, and perform multi-step research.

Gemini Agentic ReAct Strategies

Goal

Enable Gemini to solve complex tasks by combining natural language reasoning with external actions (tools), mimicking how humans operate in the real world.

The ReAct Paradigm (Reason & Act)

1. The Thought-Action Loop

• Concept: Instead of answering immediately, the model enters a loop of reasoning and acting until the answer is found.
• Workflow:
  1. Reason (Thought): The model analyzes the request and generates a plan of action.
  2. Act (Action): The model performs a specific action (e.g., Search, CodeInterpreter).
  3. Observe (Observation): The model waits for the result of that action from the external tool.
  4. Update: The model uses the new information to update its reasoning and determine the next step.

2. Example Structure

To implement ReAct, the prompt must enforce a strict format so the system knows when to stop generating and execute code.

Standard ReAct Format:

CopyQuestion: [Input]
Thought: [Reasoning about what to do next]
Action: [Tool Name]
Action Input: [Arguments for the tool]
Observation: [Result from the tool - User/System provides this]
... (Repeat until solution is found)
Final Answer: [The conclusion]