I gave Claude Code full Kubernetes cluster access and told it to fix a networking issue. It escalated through troubleshooting sessions, force-overrode etcd, and deleted netplan configurations across all control plane nodes. The cluster was gone. Forty minutes. No gate stopped it anywhere in the chain. This talk is the full incident: the actual command sequence, why the AI thought each escalation was reasonable, the moment I realized what was happening, and why the obvious fix (put a human back in the loop) is the wrong lesson. The whole point of AI agents in operations is autonomous speed. If you slow them down with manual approval on every action, you've just built an expensive autocomplete. The real question is: how do you let an agent operate fast inside a boundary it can't break out of? I spent the next six months answering that question. The result is an Eight Guardrails Framework enforced across multiple layers of the stack. I'll show you what those layers are, how they interact, where each one catches failures the others miss, and why the most important guardrail is the one most teams skip entirely. If your team is experimenting with AI agents in infrastructure, you're going to hit this wall. The question is whether you hit it in production or in this talk.
Principal Training Architect with 30 years of infrastructure experience across federal, Fortune 50, and startup environments. Has personally taught over 100,000 engineers platform engineering and AI/ML infrastructure. Workshop ranked third most-selected at KCD Texas 2026. Speaker at KubeCon EU Cloud Native University and KubeAuto AI Day Europe, Amsterdam. The cluster deletion incident described in this talk actually happened.
