OfficeBench-Pro: Benchmarking Agents for Professional Office Work

Professional Office work requires factual accuracy, analytical soundness, and aesthetic quality to jointly serve a workplace objective, yet existing benchmarks only partially capture their interdependence. We introduce , comprising expert-authored and synthesized tasks across professional domains. Agents create or revise Office deliverables from heterogeneous sources, with native structures essential to analysis or delivery. Our synthesis framework combines source-grounded design, execution-guided evolution, and independent task and rubric review. We adapt Agent-as-a-Judge to trace source evidence, test required native behavior, and visually inspect rendered outputs. Across six models under a shared execution protocol, the highest overall scores are 51.3% on expert tasks and 50.8% on synthesized tasks; no model leads all three quality dimensions in either collection. Qualitative analyses of archived submissions reveal unsupported premises propagated across deliverables, broken calculation dependencies, and visual organization that obscures analytical relationships.