AI agents in ops need limits, not just skills. Every agent gets an autonomy level, guardrails, named approvers, and a sandbox to test in before it touches a real load. Plus one switch to pause them all. What would you never let an agent do alone?