NightShift

Watch it handle an incident

A broken deploy takes checkout down. The agent is paged, investigates, names the cause and proposes a rollback; I approve it; a separate component rolls back and checks the store recovered. About a minute, or step through it at your own pace.

1 of 8

One real incident

A bad deploy on the store I built, recorded on 24 Sep 2026 and replayed exactly as it happened.

  1. The storeFour small services on AWS Lambda. A deploy breaks one of them.
  2. An alarmCloudWatch sees the errors and pages.
  3. The agentAn LLM that investigates with read-only tools and names the cause.
  4. A humanApproves one exact action, or doesn't.
  5. The ActorCarries the action out and checks that it worked.

A replay of a recorded run, not a live one: run 01-bad-deploy-20260924T023013Z on 24 Sep 2026, commit be93fc6, the agent on ministral-14b-latest. The deployed agent now runs on Gemini Flash Lite, which the benchmark found right more often. The chart is the orders service's requests and errors per minute, read from CloudWatch after the run. This page makes no call to AWS or to a model.