Skip to main content

Agent Configuration

Guide to configuring Agentrix agents through claude/config.json.

Overview​

Configuration controls agent behavior, model selection, permissions, and integrations.

config.json Schema​

{
  "model": "claude-sonnet-4.5-20250929",
  "fallbackModel": "claude-3-5-sonnet-20241022",
  "maxTurns": 50,
  "systemPrompt": {
    "path": "system_prompt.txt",
    "mode": "append"
  },
  "settings": {
    "permissionMode": "default",
    "allowedTools": ["Read", "Edit", "Grep", "Bash"]
  },
  "mcpServers": {
    "enabled": ["filesystem", "git"],
    "directory": "mcp-servers"
  },
  "skills": {
    "enabled": ["code-formatter"],
    "directory": "skills"
  }
}

Configuration Options​

Model Settings​

  • model: Primary model (default: "claude-sonnet-4.5-20250929")
  • fallbackModel: Backup model if primary unavailable
  • maxTurns: Maximum conversation turns (default: 50)

System Prompt​

  • systemPrompt.path: Relative path to prompt file
  • systemPrompt.mode: "append" (add to base) or "replace" (override base)

Permission Mode​

  • default: Requires approval for destructive operations
  • acceptEdits: Auto-approve Edit tool
  • bypassPermissions: Auto-approve all tools (use with caution)
  • plan: Show plan before execution

Tool Allowlist​

Restrict tools to specific set:

{
  "settings": {
    "allowedTools": ["Read", "Grep", "Bash"]
  }
}

Best Practices​

  1. Always set maxTurns to prevent runaway loops
  2. Use allowedTools for security-sensitive agents
  3. Keep system prompts concise and focused
  4. Test configuration changes locally