Talk
AI Agents for Kubernetes Ops: Live incident recovery with MCP

I presented “AI Agents for Kubernetes Ops: Live incident recovery with MCP” at AgentCon Helsinki. The live demo ran without any hiccups, thanks to Claude Sonnet 4 for not hallucinating too much.
The demo wired an agent built with the AWS Strands Agents SDK to a set of MCP servers and let it work a real incident on an Amazon EKS cluster end to end:
- Inject a controlled pod error
- Prometheus fires a rule to Alertmanager
- The agent debugs the cluster through the EKS MCP server and writes an incident summary
- It sends the summary by email through the SNS MCP server
- It proposes a fix through the GitHub MCP server
- After manual approval, it applies the patched Deployment via ArgoCD
The manual approval step is the part I would not remove. The interesting question is not whether an agent can produce a patch, it is where you put the human.


Thanks
Thanks to Siili Solutions for hosting, to Microsoft, and to the organisers Vesa Nopanen, Vesa Tikkanen, Anna Silvonen, Terho Antila, Vibha Deshpande and Teemu Lainiola for the opportunity.