Agents keep passing their tests while failing the work that matters
Four new benchmarks converge on one finding: outcome-based evaluation hides invalid execution. For agents that move money and write code, the middle of the trace is where trust breaks.