AI agent logs expose reproducibility gaps

AI agent logs reveal significant reproducibility gaps, where autonomous agents frequently fail even after initial successes, especially in web navigation tasks. Research, including the SWE-chat corpus, highlights that less than half of agent-produced code survives into user commits, exposing a critical discrepancy between benchmark scores and real-world reliability.