Yesterday's level: your AI agent wanted to run kubectl delete pod api-7f9c -n prod on a CrashLoopBackOff pod owned by a 3-replica Deployment. Illustrative scenario.
The answer is APPROVE. Here is why, and the trap most people miss:
1. A pod owned by a Deployment is cattle. The ReplicaSet notices the missing replica and schedules a fresh pod within seconds. Nothing is lost that the pod spec doesn't recreate.
2. Two healthy replicas keep serving while it restarts, so users see nothing.
3. THE TRAP: a PodDisruptionBudget does NOT protect you here. A PDB only limits voluntary evictions through the Eviction API (kubectl drain, node upgrades). A plain kubectl delete pod bypasses it completely. If you were counting on the PDB, check the replica count yourself first: kubectl get deploy -n prod and look at READY.
4. When the answer flips to DENY: a bare pod with no controller, a StatefulSet pod with local state, or the last ready replica. Then deleting it is an outage or data loss.
5. A restart is not a fix. If the config reload is still bad, the new pod CrashLoops again. Let the agent read the logs (kubectl logs --previous), not just delete things.
Guardrail tip: give your agent a read-only kubeconfig and let deletes go through you. The Claude Code Infra Guardrails Lab in the Classroom has a tested hook that blocks kubectl delete and drain unless you approve.
Level 3 drops tomorrow: a Terraform plan that looks clean but isn't. Would you have approved this one? One word below.