Baidu puts 200 real workflows into agent testing

Baidu has launched DuMateBench, a benchmark of 200 real-session tasks spanning eight scenarios and 17 capabilities rather than clean, single-turn prompts. The DuMateBench paper says tasks preserve workspace state and run in Docker environments with insufficient, unstable, and noisy conditions; coverage of Baidu’s launch describes the focus on usable deliverables. Builders should file this beside the case for skill retrieval as an agent advantage: offline scores are becoming less persuasive unless the harness tests the mess production workflows actually contain.