Researchers audit tool-using agent benchmarks for hidden contract violations

Academic study reveals that agent benchmarks assume tools work as advertised without verification, finding systematic defects in every rerun—critical for accounting use cases where agents execute r...