ReAct Pattern

Translation in Progress

This page is being translated. Content below is a placeholder.

What is ReAct

ReAct = Reasoning + Acting

A paradigm that combines reasoning and acting for AI Agents.

The Three-Step Loop

1. Thought

The LLM analyzes the current situation:

• What has been done?
• What information is needed?
• What's the next step?

2. Action

Choose and execute an action:

• Call a tool
• Generate output
• Request more information

3. Observation

Observe action results:

• Tool execution result
• Error messages
• New information discovered

Example

User: "What's the total line count of all TypeScript files in the project?"

Thought: Need to find all .ts files and count lines
Action: run_command("find . -name '*.ts' | xargs wc -l")
Observation: Total 12345 lines

Thought: Got the result, can answer now
Action: Output answer
Result: "This project has 12,345 lines of TypeScript code"

Implementation in gemini-cli

// Simplified ReAct loop
async function agentLoop(task: string) {
  let context = task

  while (true) {
    // Thought: LLM decides what to do
    const response = await llm.generate(context)

    // Check if task is complete
    if (!response.hasToolCall()) {
      return response.text()
    }

    // Action: Execute tool
    const result = await executeTool(response.toolCall)

    // Observation: Add result to context
    context = addToContext(context, result)
  }
}

Why ReAct Works

1. Decomposition: Complex tasks → simple steps
2. Feedback: Each step has observable results
3. Correction: Can adjust strategy based on results
4. Transparency: Reasoning process is visible

Summary

• ReAct = Reasoning + Acting
• Core loop: Thought → Action → Observation
• Well-suited for complex task decomposition and execution

Next

Understand how to interact with LLM APIs: LLM API Interaction →