From Automated Program Repair to Automated Agent Repair: Why Failed Trajectories Matter
Xiangxin Zhao
ASE 2026 Student Research Competition (SRC), Munich, Germany · Accepted
Abstract. Automated Program Repair (APR) has shown that buggy programs can often be repaired rather than rewritten. As LLM coding agents increasingly perform software engineering tasks, an analogous question arises: can failed agent executions be repaired instead of restarted? We answer the prerequisite question with, to the best of our knowledge, the first execution-grounded audit of failed coding-agent trajectories. Our preliminary results show that failed trajectories retain substantial recoverable value, completing a median of 75% of task subgoals (n = 27), while none of the audited agents ever states the correct fix (0/26). These findings establish the empirical foundation for Trajectory-Centric Automated Agent Repair (TC-AAR), extending APR from programs to agent executions.
Recommended citation: Xiangxin Zhao. 2026. "From Automated Program Repair to Automated Agent Repair: Why Failed Trajectories Matter." In Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE '26), Student Research Competition. ACM. doi:10.1145/3832783.3844579
Download Paper | Download Bibtex