Agent Mode★

Agent mode is the configuration setting that dictates the role, permissions, and tool access scope of an agent for a specific session. Instead of running a single general-purpose prompt, developers change the agent mode depending on the phase of the development lifecycle.

You interact with agent modes when:

• Switching your IDE assistant from "chat mode" to "refactor mode."
• Configuring an agent to act as a read-only code reviewer instead of a writer.
• Enabling a data-science agent to execute local scripts in a sandboxed interpreter.

Standard Agent Modes

AI agent systems group instructions and tool accessibility into distinct modes:

1. Architect Mode:
   • System Prompt: "Focus on planning, design, and structural feedback. Do not edit files or write code."
   • Tools: Read-only (view files, list directories, search web).
2. Builder Mode:
   • System Prompt: "Implement the approved designs. Edit code, create files, and compile projects."
   • Tools: Full-write access (replace content, create file, run build commands).
3. Code-Interpreter Mode:
   • System Prompt: "Analyze datasets. Write Python code and execute it inside the sandbox."
   • Tools: Sandbox execution (run script, print data logs).

Field Applications

1. Fullstack Developers (Routing Tool Calls by Mode)

Developers write routing logic to disable or enable tool schemas in the model request based on the selected mode:

• Code Example:

  function getToolsForMode(agentMode) {
    const readTools = ["view_file", "list_dir"];
    const writeTools = ["write_to_file", "replace_file_content"];
    const executeTools = ["run_command"];
    
    switch (agentMode) {
      case "architect":
        return readTools; // Read-only
      case "builder":
        return [...readTools, ...writeTools, ...executeTools]; // Full access
      case "interpreter":
        return ["execute_python_script"]; // Isolate to Python execution
      default:
        return readTools;
    }
  }