Agentic AI meets Kubernetes Ops

Tech talk and demo inspired by the Kubecon talk by https://github.com/janakiramm

Transcript

1. Agentic AI meets Kubernetes Ops Self-healing operations with AI agents, MCP and A2A Gopal Ramachandran June 2026
2. It's 2:07 AM. Your pager just went off.
3. $ kubectl get pods NAME READY STATUS RESTARTS api-7d4f9c-x2lp 0/1 CrashLoopBackOff 7 api-7d4f9c-9k2m 0/1 OOMKilled 6 api-7d4f9c-q81v 0/1 CrashLoopBackOff 7 $ ▌
4. • Alert fatigue • Manual toil • Fragmented tooling The Problem What if agents could handle this autonomously?
5. Building blocks MCP One protocol for all tools A2A One protocol for agent comms Agentic AI LLMs that reason and act
6. Architecture Orchestrator Python 1 · Diagnose CrewAI + GPT Port 10001 2 · Remediate Google ADK + Gemini Port 10002 3 · Verify Strands + Claude Port 10003 Kubernetes MCP server Kubernetes cluster
7. Considerations • Not everything needs to be an agent • If we got MCP then why A2A?
8. Demo We will break a deployment on purpose and watch the agents fix it.
9. Incident summary T+0s OOMKill detected Watch API event fires T+30s Root cause named 64Mi limit vs 256Mi demand T+60s Deployment recreated memory limit → 512Mi T+90s Verified healthy 3/3 pods Running, 0 restarts
10. Next steps • Make it sovereign • Needs guardrails for production deployment • Memory of past incidents • Think about more use cases!