The paper frames root-cause attribution for long-horizon AI agent failures as a search problem rather than a single-pass classification task, since relevant diagnostic evidence in execution logs is often sparse and disconnected from the visible failure point. The authors introduce Continual Search, an iterative framework that repeatedly prompts an LLM judge across multiple turns to keep searching for unresolved diagnostic evidence rather than stopping after one pass. Evaluated on four existing benchmarks plus a new MegaRCA-Mix dataset of 50 human-annotated long-horizon failure trials, it improves GPT-5.5’s F1 score from 0.349 to 0.498, a gain of more than 40%. The results indicate that more effective search can outperform simply scaling up model size for this diagnostic task.
