AI / LLM / Agents
SRE Incident Response Environment
An OpenEnv-compatible environment that trains AI agents on production incidents — alert triage, root cause analysis, and cascading failure resolution.
01 / Problem
The problem
Production systems fail in interconnected ways, but there is no standard environment for training or evaluating AI agents on real incident response — existing benchmarks test reasoning in isolation, never under an active outage.
02 / Response
The solution
Built a realistic incident environment exposing logs, metrics, traces, and service dependency maps, with three progressive scenarios and quantitative scoring across classification, investigation, diagnosis, remediation, and communication.
03 / Product
Key capabilities
Three progressive incident scenarios
Full observability surface — logs, metrics, traces, dependency maps
Five-axis quantitative scoring
OpenEnv-compatible interface
Containerised, reproducible runs
Scenarios modelled on real production outages
04 / Outcome
Impact
05 / Reflection
What I learned
Designing the environment is harder than solving it — the scoring rubric is where the real judgement lives, and a poorly shaped reward teaches the wrong behaviour.