My AI-agent waste detector scored zero false positives. Then I ran it on a real trace.
The developer of an AI agent waste detection tool, Clew, initially hypothesized that structural cycles and embedding decay could predict multi-agent failures. However, testing on the MAST-Data dataset yielded poor results, indicating the metric was largely measuring trace length rather than failure. The project pivoted to detecting redundant loops and handoffs that consume tokens, inspired by a paper suggesting a cascaded approach of structure then semantics for cycle detection. Rigorous self-validation methods were implemented, including pre-registered GO/KILL criteria and code that prevents accidental data leakage, to ensure honest results. AI
IMPACT This tool aims to reduce operational costs for multi-agent AI systems by identifying and eliminating token waste.