Han Li, Yifan Yao, Letian Zhu, Rili Feng, Hongyi Ye, Jiaming Wang, Yancheng He, Pengyu Zou, Lehan Zhang, Xinping Lei, Haoyang Huang, Ken Deng, Ming Sun, Zhaoxiang Zhang, He Ye, Jiaheng Liu

Researchers have introduced CodeTracer, a tracing architecture designed to debug code-generating AI agents by reconstructing their full state transition history and pinpointing where failures originate. The system parses execution artifacts from multiple agent frameworks, builds hierarchical trace trees with persistent memory, and performs failure onset localization to identify error chains. The team also released CodeTraceBench, a benchmark dataset with stage and step-level supervision from four major code agent frameworks tested on diverse tasks like bug fixing and refactoring. Experiments show CodeTracer substantially outperforms simpler baselines and can recover originally failed runs when diagnostic signals are replayed.