The Agent Waste Index has scored 341,054 agent trajectories from 11 public datasets to identify inefficiencies such as loops, blind retries, oversized tool output, and abandoned runs. The analysis reveals that loop and retry behaviors are prevalent in weaker models like Llama agents but rare in Claude 3.7 Sonnet and Qwen3-Coder-480B.

  • Oversized tool output varies significantly by scaffold, with OpenHands and SWE-agent carrying >20k-character observations in over half of runs.
  • The longest fifth of runs by step count consumes 40% of estimated spend while resolving at a lower rate than the shortest fifth.
  • The project provides a pipeline and detectors to rebuild the Index from the hub, with tables available online for community verification.

The authors provide correction mechanisms for dataset publishers and aim to improve transparency in coding agent performance.