Fuzzy Glade: Evidence-Grounded, Agent-Guided Analysis of Agent Trajectories at Scale

As agents get more capable, model training requires richer behavioral signal than what static evaluation metrics like pass rate offer. Agent trajectories are a gold mine for such insights, but their size, volume, and granularity make them difficult to analyze for high-quality insights. We introduce Fuzzy Glade (FG)—a comprehensive framework and application for performing agent-guided analysis of agent trajectories at scale in a reproducible, evidence-annotated, and easy to review manner. We also present a case study from a real use of FG in the model training of our production model—detecting and profiling reward-hacking behavior in evaluation and reinforcement-learning rollouts. We release the reward-hacking dataset sourced from this use case alongside human annotated ground truth as an evaluation of the tool.