PreAct: Computer-Using Agents that Get Faster on Repeated Tasks

The article introduces PreAct, a novel framework for computer-using agents that enables them to accelerate repeated task execution by compiling successful runs into state-machine programs, achieving speedups of 8.5-13x without further language model calls. PreAct incorporates a verification mechanism to ensure that the screen state matches expectations before replaying actions, and it employs a store-time check to filter out faulty programs, resulting in improved task performance across mobile, desktop, and web benchmarks. This approach is significant for practitioners as it enhances efficiency in task automation by allowing agents to learn from past experiences and reduce the computational overhead typically associated with repeated tasks.