Localizing Emergent Failures in Agentic AI: Recovering Minimal Repair Families via Counterfactual Replay
Graph-Constrained Joint Replay (GCJR) is proposed, which slices failure-relevant events from an execution dependency graph, constructs graph-feasible singleton and pair candidates, and verifies them by replay with paired clean counterparts, and is exact within its declared graph domain.
Bingbing Li, Yumeng Song, Zhongming Yao et al.
· 0 citations