AI agent benchmarks are misleading, study warns

A study by Princeton University shows that benchmarks made for AI agents don't account for costs and are prone to overfitting.