LangChain's Deployed Engineering team built an autonomous SRE agent to reduce Kubernetes triage and remediation toil. The system uses a multi-agent architecture with specialized subagents (pod-inspector, scaling-analyzer, log-analyzer, etc.) running in parallel, with Claude Sonnet for orchestration and Claude Haiku for cheaper scheduled health checks. A strict read/write split enforces safety: the agent can read the entire cluster autonomously but all writes require human approval via Slack. LangSmith tracing was central to debugging — it revealed a 20-call scheduled check wasting tokens, a runaway tool loop, and false-positive CRITICAL alerts on single-replica services. Human edits to proposed changes are captured as labeled training data. LangSmith Engine further automates the improvement loop by clustering traces into issues, writing prompt/code fixes as GitHub PRs, and suggesting regression tests. The project is open source.
