How Do AI Agents Waste Your Money?

Researchers introduce a unified framework to measure and identify token waste across heterogeneous AI agent systems by recording trajectories, identifying removable spans using an LLM-as-a-Judge, and validating them through direct interventions. In online pruning experiments, the authors demonstrate that removing 41 to 46 percent of the model's reasoning preserves task performance across two agent benchmarks. Their analysis reveals that token waste varies more across tasks and agent settings than across underlying models. Open article →