How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope
English summary
This paper studies the transition from conversational AI assistants to autonomous agents using production data from Perplexity's Search and Computer products. It finds that the Computer agent performs 26 minutes of autonomous work per user session, compared to 33 seconds for Search, with 55% lower dissatisfaction rates. The agent reduces task completion time from 269 to 36 minutes and lowers estimated cost by 94% compared to humans using Search alone. Additionally, it expands the scope of work to tasks that cross occupational boundaries, require higher-order cognition, and bundle interdependent subtasks, unlocking work activities absent in Search usage.
Chinese summary
本文利用Perplexity的Search和Computer产品的生产数据，研究从对话式AI助手到自主代理的转变。研究发现，Computer代理每个用户会话执行26分钟的自主工作，而Search仅33秒，不满率降低55%。与仅使用Search的人类相比，代理将完成任务的时间从269分钟减少到36分钟，成本降低94%。此外，代理还扩大了工作范围，包括跨职业边界、需要更高认知和捆绑相互依赖子任务的任务，解锁了Search使用中不存在的活动。
Key points
Computer agent performs 26 minutes of autonomous work per session versus 33 seconds for Search.
Computer代理每个会话执行26分钟的自主工作，而Search仅33秒。
Dissatisfaction rates are 55% lower on Computer than on Search.
Computer的不满率比Search低55%。
Completion time reduced from 269 to 36 minutes; cost reduced by 87-94% compared to humans with Search.
完成任务时间从269分钟减少到36分钟；与仅使用Search的人类相比，成本降低87-94%。
Computer expands work scope to include higher-order cognition, cross-boundary tasks, and composite tasks.
Computer将工作范围扩大到包括更高阶认知、跨边界任务和复合任务。