Holo4: powering generalist computer-use agents

OSWorld-Pro introduces a process-based evaluation for Computer-Use Agents, addressing the limitations of functional verifiers used in OSWorld. While traditional assessments focus on final deliverables, OSWorld-Pro provides transparency into how and why agents fail. By utilizing over 300 tasks and 2800 subgoals grounded in 67,000 human annotations, it distinguishes between keyboard input errors and graphical UI click failures. This procedural approach offers critical insights for mitigation strategies, while Holo4 aims to power generalist computer-use agents.