“On the OSWorld 2.0 benchmark, current AI agents commonly fail by losing track of constraints, missing information that arrives mid-task, guessing rather than asking the user, and skipping verification.”