尽管2024-2025年主要AI实验室纷纷推出计算机使用原型并在基准测试中取得亮眼成绩，但这些AI代理在真实应用场景中的表现却远低于预期。[1]在OSWorld-V2基准测试中，最佳模型的任务完成率仅为20.6%，Agents' Last Exam的最佳结果也仅有26.2%。[1]相比之下，基准测试本身显示的完成率从60.76%到97.4%不等，而人类在WebGames基准上的成功率超过95%。[1]
问题的根源在于AI代理的工作方式与人类存在根本差异。[1]GPT-5.5和Claude Opus等模型倾向于通过API调用的方式绕过用户界面，而非像人类一样直接操作界面完成任务。[1]这种策略导致计算预算的低效配置，视觉定位、低级操作和工具使用的开销占据了代理计算预算的主要部分。[1]仅通过扩大模型规模无法解决这一核心问题，真正的突破需要改变代理与计算机界面的交互方式。[1]
Recent analysis reveals a significant gap between how well AI agents perform on standardized benchmarks and their actual effectiveness in real-world computer use scenarios. While major AI laboratories have released computer-use prototypes in 2024-2025 with impressive benchmark results ranging from 60.76% to 97.4% completion rates, the performance drops dramatically in practical applications.[1] On the OSWorld-V2 benchmark, the best-performing model achieved only a 20.6% completion rate, while another leading benchmark, Agents' Last Exam, reached just 26.2%.[1]
The underlying issue stems not from model size constraints but from how AI agents approach interface interaction.[1] Rather than directly manipulating user interfaces as humans would, leading models like GPT-5.5 and Claude Opus tend to circumvent interfaces by relying on API calls instead.[1] This approach introduces computational inefficiency and reduces reliability. Research into OSWorld 2.0 data shows that agents' computational budgets are dominated by expenses for visual positioning, low-level operations, and tool usage, suggesting fundamental architectural misalignment with human interaction patterns.[1] Despite human success rates exceeding 95% on benchmarks like WebGames, AI models remain significantly below this threshold.[1]