SREGym open-sources a benchmark replaying real incidents like Cloudflare and Reddit outages for SRE agents
tianyin_xu · x · 2026-11-19
SREGym is an open-source benchmarking platform (419 stars, MIT) that builds live, reproducible system environments for evaluating SRE agents. It already simulates well-known real-world cloud failures — each with a linked postmortem and simulation: Cloudflare WAF regex CPU exhaustion, admission webhook TLS mismatch, conntrack table exhaustion, GKE IP exhaustion, Kafka poison pill, and the Reddit Pi-Day outage. The thread grew out of a call for open, independently verifiable benchmarks, since internal evals tend to mirror a team's own blind spots.
More from coding & agent
- OpenCode fix cuts GPT-6.2 uncached input per step from ~14,000 tokens to 3 — victor_explore · 2026-11-19
- Remotion ships real motion blur, addable via AI agents with one prompt — Vjeux · 2026-11-19
- An LLM is not an agent: the 10 building blocks of agentic systems, from orchestration to RAG — goyalshaliniuk · 2026-11-19
- Uber, SpaceX and Nvidia engineers pack first AI Software Factory summit PortCon — Pavan_Belagatti · 2026-11-19
- Dev builds James Bond-style stealth game with GPT 6.2 Astra — chongdashu · 2026-11-19
- Why autonomous agents outgrow MCP and collapse back to APIs + scoped tokens — uriwa · 2026-11-19