An AI agent’s control loop plans and executes steps toward a goal, maintains context, and applies configured policies and permissions. Tool use can be part of the loop: the controller checks a tool, runs it, and returns its results for further reasoning or action. section 2.6.1 section 2.6.2
Build operational safeguards into tool use:
- Check that each tool is available, permitted, and within policy before execution. section 2.6.2
- Document each tool’s purpose, access scope, permissions, and human checkpoints; require approval and maintain audit records for sensitive actions. section 2.6.2
Security agent
A security agent can trigger a vulnerability scan, retrieve and parse results, compare findings with policy, open tickets for high-severity issues, and post a summary for incident response. section 2.6.2

