Microsoft Research AgentRx: Open-Source Framework Achieves +23.6% Accuracy Localizing Exact Failure Step in Agent Trajectories

Microsoft Research·medium signal

Released March 12, 2026: AgentRx automatically diagnoses why agents fail by synthesizing executable constraints from tool schemas and domain policies, then validating each step in the execution trace to identify the first unrecoverable point. Outperforms prompting baselines by +23.6% in failure localization and +22.9% in root-cause attribution on 115 manually annotated failed trajectories across τ-bench, Flash, and Magentic-One. Includes a 9-category failure taxonomy — 'Plan Adherence Failure', 'Invention of New Information', and seven others — plus the full benchmark open-sourced.