https://t.co/7xQOYbt3R7

LangChain built an AI agent that diagnoses problems in its Kubernetes clusters on its own but cannot change anything without a person's approval. Kubernetes is the system that runs a company's services across a cluster of machines, and it reports a constant stream of status signals, from container restarts to autoscaler states, with almost no synthesis. In an article republished in LangChain's post, Deployed Engineer Eric Johanson writes that 90% of on-call triage is mechanical work that mostly comes back clean, and that this toil trains engineers to skim past important alerts. The agent is meant to take over that triage, propose fixes, and pull in a human only when the infrastructure itself has to change.

The agent can read the entire cluster, while every write action, such as scaling a deployment, restarting a rollout or patching an autoscaler, lives inside a single change-executor subagent. Each write tool pauses for a human, who approves, rejects or edits the proposed change from a Slack message. Johanson says the split is enforced in the code rather than by instructions to the model: read and write tools sit in separate modules, the orchestrating agent cannot reach a write tool at all, and the cluster's own access controls mirror the same split, granting cluster-wide read access and tightly scoped write access. The team also withheld broad tools such as a helm upgrade, which can rewrite dozens of resources the approver cannot see from the prompt. Scaling a deployment to 3 instances can be judged at a glance, and Johanson argues the approval step protects production only when the person can judge what they are approving.

Cost shaped the rest of the design. For an investigation, an orchestrator running on Claude Sonnet sends specialist subagents for pods, scaling, logs, security and other areas to read the cluster in parallel, then combines their findings into one prioritized report; those read-only subagents and the scheduled health check run on Claude Haiku. The scheduled check originally ran the full orchestrator, about 20 model calls, just to confirm that everything was healthy. Per-run cost breakdowns in LangSmith, LangChain's tracing service, led the team to collect cluster state in plain Python and make a single Haiku call instead, which Johanson reports cut the cost of each check by 95 to 99% with no loss in catching issues. That figure and the reliability claims are LangChain's account of its own system; the article publishes no measurements behind them.

The same traces feed the agent's tests. Each human approval, edit or rejection is recorded as feedback on the run, and mistakes, such as a subagent repeatedly flagging intentionally single-replica services as critical, become test cases that a prompt or model change must pass before it merges. LangChain says it has begun rolling the agent out to some LangSmith customers, and the code is published in the sre-agent repository.