Self-Healing Kubernetes Gets Real — and Risky: Running AI Agents on Amazon EKS
AI agents capable of autonomously diagnosing and remediating Kubernetes cluster issues are now production-ready on Amazon EKS, powered by the EKS MCP server, AWS DevOps Agent, CloudWatch agentic investigations, and Amazon Bedrock AgentCore. The post details how agents interact with clusters via IAM-scoped MCP tools (read-only by default), documents real-world incidents where over-permissioned agents caused production data loss, and proposes a four-phase maturity model — from read-only observation to governed autonomy. Key safety guidance emphasizes least-privilege IAM roles, Cedar-based policy enforcement, admission control via Kyverno/OPA, human approval for destructive actions, and full audit trails. The core argument: the capability gap for agentic Kubernetes ops has closed, but the governance gap has not, and guardrails must live outside the agent's reasoning loop.