What is AI Agent

Translation in Progress

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

Definition

An AI Agent is not just an LLM (Large Language Model), but a system that can autonomously complete tasks.

Core formula:

Agent = LLM + Tools + Loop

• LLM: The "brain" for reasoning and decision-making
• Tools: Capabilities like reading files, executing commands, web search
• Loop: Iteratively call LLM until task completion

Agent vs Chatbot

Feature	Chatbot	Agent
Interaction	Q&A	Task execution
Capabilities	Text generation only	Use tools
Autonomy	Requires each instruction	Can plan and execute
State	Stateless or simple context	Complex state management

Agent Capabilities

What can an Agent do?

1. File operations: Read, write, edit code
2. Command execution: Run shell commands
3. Web interaction: Search, fetch pages
4. Complex tasks: Multi-step reasoning and execution

Example: File Modification Task

User: "Fix the bug in src/utils.ts line 42"

Agent thinking:
1. Read src/utils.ts → understand code
2. Locate line 42 → identify issue
3. Decide fix → generate patch
4. Edit file → apply changes
5. Verify → check result

How gemini-cli Implements Agents

gemini-cli's Agent implementation is in packages/core/src/core/client.ts:

// Simplified structure
class GeminiClient {
  // Main loop
  async run(userMessage: string) {
    while (true) {
      // 1. Send to LLM
      const response = await this.chat.send(message)

      // 2. Check for tool calls
      if (response.hasToolCall()) {
        // 3. Execute tool
        const result = await this.executeTool(response.toolCall)
        // 4. Feed result back to LLM
        message = result
      } else {
        // 5. Done
        return response.text()
      }
    }
  }
}

Summary

• Agent = LLM + Tools + Loop
• Differs from chatbots by having tools and autonomy
• Core is the "think → act → observe" loop

Next

Learn more about the pattern behind Agent reasoning: ReAct Pattern →