PreAct Accelerates Computer-Using Agents on Repetitive Tasks

Computer-using agents, which interact with software by simulating clicks and typing, typically process every task as if it were new. This means that even when asked to repeat a task, the agent re-analyzes the screen and re-reasons each action, incurring the full computational cost every time. This approach limits efficiency for repetitive operations. A new system called PreAct addresses this by allowing agents to accelerate on tasks they have previously completed. When an agent successfully performs a task for the first time, PreAct compiles that entire execution into a compact state-machine program. This program consists of states that verify screen conditions and transitions that execute actions. For subsequent runs of the same task, PreAct directly replays this compiled program, leading to an 8.5 to 13 times speed improvement by eliminating per-step language model calls. Crucially, this replay mechanism is not "blind"; at each step, PreAct verifies that the screen state matches expectations. If a discrepancy is detected, control is immediately returned to the agent for dynamic problem-solving, ensuring robustness. A rigorous validation process ensures that only correctly completed task programs are stored for reuse, preventing the accumulation of faulty programs.