Troubleshooting Like a Senior on Day 1: ReAct Agents with Real-Time Cluster Evidence

If a production incident hits on your first day, can you debug it? Or if you are a senior engineer, do you find it impossible to download your years of debugging intuition into a new hire’s head?
Kubernetes troubleshooting often depends on undocumented decision paths: where to look first, which signals to trust, and how to turn a sea of logs into a testable hypothesis.

In this talk, we introduce KUBE-RCA, an open-source incident assistant that plugs into your preferred external LLM and provides real-time cluster evidence (metrics, logs, events) as structured context. Using a ReAct loop, the agent proposes hypotheses, runs an allowlisted set of read-only commands/queries, ties each claim back to evidence, and publishes a concise RCA draft directly to your team channel.

We’ll share the design decisions behind our guardrailed execution loop and how we encode SRE intuition into prompts and checks. You’ll walk away with an understanding of how to make incident response more systematic. So engineers of any tenure can resolve issues faster, with less senior interruption.