Analysis of 80,000 AI agent sessions reveals that nearly 9 in 10 loop attempts fail, with a predictive accuracy of 0.814 AUC.

Study analyzed 80,000 AI agent sessions to identify failure patterns and performance metrics

88.7% of loop operations failed across the analyzed agent sessions

Model achieved an AUC (Area Under the Curve) score of 0.814, indicating reasonable predictive performance

Research published as CAUM Agent Waste Report 2026, suggesting a focus on improving agent efficiency and loop reliability