The problem
When a pod crash-loops or a service degrades, on-call engineers lose time manually correlating Kubernetes resource changes, logs, and runbooks before they can even start fixing anything.
The approach
I built a PydanticAI investigation agent with five read-only tools that pulls Kubernetes resource history and deduplicated logs against Elasticsearch-indexed evidence, forces every claim to cite a concrete source, and separates confirmed observations from hypotheses so it never asserts something it didn't actually see.
The outcome
Validated with 22 automated tests across three failure scenarios and a 30-case benchmark tracking retrieval quality, latency, and cost.
More detail on this project — the agent's tool design, example investigations, and what I'd change next time —
is coming soon.