people · May 15, 2026

Microsoft Researchers Introduce AgentLens Framework for SWE-Agent Trajectory Assessment

Microsoft researchers Priyam Sahoo, Gaurav Mittal, Xiaomin Li, Shengjie Ma, Benjamin Steenhoek, and Pingping Lin released AgentLens on May 13, 2026, to evaluate SWE-agent processes beyond binary pass rates. AgentLens merges multiple passing solutions into Prefix Tree Acceptor references and labels actions across exploration, implementation, verification, and orchestration stages using trajectory history. The framework accompanies AgentLens-Bench, a dataset of 1,815 annotated trajectories from SWE-bench tasks under OpenHands scaffold. Among passing trajectories, AgentLens identifies 10.7% as Lucky Passes involving regression cycles or blind retries. Microsoft released the AgentLens SDK and dataset on GitHub on May 14, 2026.