Permission Request★

A permission request is a security gate in an agent's execution loop. When the model requests a sensitive action (like running a terminal command or writing to a file), the client harness pauses execution and prompts you to approve, modify, or reject the tool call.

You trigger permission requests when:

• Running terminal agents that ask Run command 'git push'? [y/N].
• IDE assistants ask for permission to write modifications to a file.
• Security configurations block network calls by default.

Balancing Autonomy and Safety

Permission requests solve the security issues of agent systems. If an agent loops autonomously, it could execute destructive actions. By introducing a Human-in-the-loop (HITL) checkpoint, you limit the blast radius:

[Tool Call: delete_file] ──► Harness Pauses Loop ──► Console Prompt: "Approve?" ──► [User Types 'Y'] ──► Execution Completed

However, permissions introduce interruptions. If an agent asks for approval on every read-only command (e.g. listing directories or reading files), it wastes your time and slows development.

Field Applications

1. Fullstack Developers (Implementing Approval Gates)

Developers configure permission gates in their tool execution routing scripts, bypassing prompts for safe tools (reads) and prompting for unsafe tools (writes/executes):

• Code Example:

  async function executeToolWithPermission(toolCall) {
    const isSensitive = ["run_command", "write_file", "delete_file"].includes(toolCall.name);
    
    if (isSensitive) {
      // Pause loop and prompt developer via terminal CLI
      const approved = await promptUserForApproval(toolCall);
      if (!approved) {
        return "Error: Action rejected by the user.";
      }
    }
    
    // Execute function locally
    return runLocalFunction(toolCall.name, toolCall.args);
  }