Cheap agent drives, GPT-6 reviews: SRE fault pass rate jumps from 78.1% to 89.5%
tianyin_xu · x · 2026-10-22
SREGym let TypeSafe Jev run SRE investigations autonomously, with GPT-6 Sol reviewing root causes. Across 21 SREGym-Lite faults (105 trials), review lifted pass rate from 78.1% to 89.5%. The key design choice: the reviewer sees only evidence Jev collected and can keep or swap candidate causes, but cannot run tools itself — the expensive model never drives the loop. Cost per diagnosis doubles from $0.017 to $0.033, and mean time rises from 31.9s to 38.4s. Notably, GPT-6 Luna Decisions in the driver's seat gains even more from review (79.0% to 95.2%), showing Jev isn't the strongest driver.
Related event: SREGym: Decision Model Plus LLM Reviewer Boosts SRE Diagnosis to 89.5%(2 posts)→
More from coding & agent
- Agent Failure Meme: 'It Was the Containers, Don't Shoot the Messenger' — rakyll · 2026-10-22
- Incident agent asks "should we pull someone else in?" during a high-severity outage — sloppenheimer · 2026-10-22
- NVIDIA's sub-1B Nemotron 3.5 Parse tops ParseBench as best open-weight doc parser — llama_index · 2026-10-22
- Chatlet: Generative UI That Emerges From Group Conversations — round · 2026-10-22
- AI Agent Nearly Botches a Redeploy, Blames Containers Instead — rakyll · 2026-10-22
- NJU's Socratic Skill Learning hits 0.801 utility, beating distillation baselines — NJU-LINK · 2026-10-22