Root-cause attribution fails on long traces because one-shot judges settle early.

Automated root-cause analysis over agent logs degrades as traces grow, because relevant evidence is sparse, spread across distant actions, and disconnected from the visible failure (arXiv 2609.13463). That makes diagnosis a search problem, and existing methods use one-shot LLM judgments that lock onto a plausible diagnosis early. Continual Search nudges the judge across successive turns to keep hunting unresolved evidence, evaluated on four existing benchmarks plus MegaRCA-Mix, 50 human-annotated failure trials on long-horizon execution-heavy tasks.