"I Ran 14 Agent Failure Scenarios Through a Guard Layer (With Cost Data From Every Run)"

How do you know if your agent is looping or actually working? I've been stress-testing AI agents...