OpenSRE ships an RL environment for incident response

OpenSRE treats a production outage as a training exercise for reinforcement-learning agents, per its GitHub repo. The Apache-2.0 toolkit calls itself an open reinforcement learning environment for agentic infrastructure incident response, complete with end-to-end tests and synthetic incident simulations.

The pitch starts with a gap. Coding agents have SWE-bench to measure themselves against; distributed production failures, the kind that page an on-call engineer at 3 a.m., have had no equivalent training set.

OpenSRE's answer is to build one. It correlates logs, metrics, traces and runbooks for root-cause analysis, links the evidence an investigation turns up, lets investigations resume where they left off, and tracks the cost of each session.

Remediation is optional, and alerts route to Slack, PagerDuty or Telegram. The repo lists 60-plus integrations spanning Kubernetes, AWS, GCP and Azure.

It's still public alpha, at 2,951 commits. The repo doesn't say how its RL reward signal is validated beyond synthetic incident simulations, so it's unclear whether an agent trained here holds up against a real outage with a real blast radius.

That gap is the whole story. OpenSRE is infrastructure for a benchmark, not a benchmark itself. I'd bet it stays a curiosity until someone runs an agent through it on incidents that aren't synthetic and publishes what happened.

At 9,345 stars and 2,951 commits, there's already an audience betting it gets there.