Simple AI setups beat purpose-built agents at finding the root cause of outages

On 90 validated failure cases from eight public datasets, three simple setups that let a model investigate freely, including OpenAI's Codex coding agent, found the root cause 65% to 77% of the time on mid-tier models. Five published root-cause agents with hand-built procedures scored 16% to 53%.

Photo: a_verdoy / Wikimedia Commons, CC BY 2.0 (cropped)

Sources

• Research arXiv（査読前） (2026-10-26)
  Free to Look, Free to Reason: Simple LLM Harnesses Outdiagnose Hand-Crafted Root Cause Analysis Agents

Editorial policy ・ Terms of use