PlatformPilot
Overview
PlatformPilot is an agentic assistant for Kubernetes alerts. It investigates through read-only tools over pod logs, events, deployment config and CloudWatch metrics, then proposes a root cause, evidence and a fix, and waits for an engineer to approve before anything is applied. The agent can run on Claude through tool use, or on a deterministic planner for demos and CI, and everything runs locally without a cloud account.
Safe by design
The agent only has read tools. The one write path, applying a fix, is reachable only through the approval endpoint, so nothing changes on a cluster without a person saying yes.
Evaluated on 40 injected failures
I wrote 40 failure scenarios across 8 classes: missing environment variable, OOM kill, bad image tag, missing secret, bad config value, misconfigured readiness probe, unschedulable pods, and a dependency with a full disk. Two of every five variants are noisy or misleading so diagnosis is not a keyword lookup. The evaluation scores root-cause accuracy, fix accuracy and whether the simulated cluster is healthy after the approved fix.
Honest limits
The scenarios are simulated snapshots served through the same tool interface a real cluster would use, and the offline planner was built against them, so its perfect score is an upper bound rather than a generalization estimate. The time saved is modeled from stated assumptions (4 minutes to triage plus 2.5 minutes per lookup), not measured on real incidents. The Terraform for a test EKS cluster and the live kubectl path are included but have not been exercised against AWS.
Observability
The app exports run counts, latency, tool calls and evaluation scores as Prometheus metrics, with a provisioned Grafana dashboard you can start with one docker compose command.
Stack
Python · FastAPI · LLM tool use · Kubernetes · Terraform · Prometheus · Grafana