OSWorld 2.0 evaluates long-horizon AI agents

Current computer-use AI agents are still far from ready for autonomous workspace execution, as long-horizon tasks expose their tendency to lose context and fail at self-verification.

* The transition from short-term tasks (under 30 steps) to multi-hour workflows (300+ steps) highlights that state tracking and error recovery remain major bottlenecks for frontier LLMs.

* The poor success rate of top-tier models (maxing out at 20.6% binary completion) suggests that current agent architectures struggle heavily with dynamic, streaming information that changes mid-task.

* Fine-grained final-state checking in OSWorld 2.0 will force the research community to move past simplistic pass/fail evaluation and focus on partial progress scoring and state validation.