LLM agents generally perform worse when required to produce a Python script for root-cause prediction than when submitting predictions directly

agents generally perform worse when required to produce a Python script that maps each time-series sample to a predicted root-cause label than when they submit predictions directly

http://arxiv.org/abs/2608.27182v1