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

A new system called PreAct allows computer-using agents to accelerate repeated tasks by compiling successful runs into small state-machine programs, according to a paper posted to arXiv. The approach replays those programs directly, bypassing the need for per-step language-model calls. The first time an agent succeeds at a task, PreAct compiles the run into a program of states that check the screen and transitions that act. On later runs it replays that program 8.5-13x faster, with no per-step language-model calls [1]. Replay is not blind: at each step PreAct verifies that the screen matches what the program expects before acting, and hands control back to the agent the moment something is off [1]. A freshly compiled program enters the store only if, re-run from a clean state, an independent evaluator confirms it solved the task. This store-time check catches programs that replay to their last step yet leave the task undone [1]. Across a mobile, a desktop, and a web benchmark, the check separated repeated runs that improve from ones that degrade as faulty programs accumulate, worth 1.75-2.6 tasks per benchmark, the same direction on all three [1]. A fallback that explores afresh when no program fits brings PreAct level with a strong record-and-replay baseline [1]. The paper also reports what did not matter: prompt wording, runtime guardrails, and whether a language model or a plain embedding retriever selects which program to reuse [1]. The work appears on arXiv, a preprint repository that has collaborated with Hugging Face to embed interactive machine-learning demos directly alongside papers, allowing readers to try models without writing code [8][9]. PreAct’s focus on making agents more efficient on repeated tasks intersects with broader discussions in AI alignment, a subfield of AI safety concerned with steering systems toward intended goals [3]. Alignment researchers study challenges such as instilling complex values, developing honest AI, and preventing emergent behaviors like power-seeking [3]. The ethics of artificial intelligence also covers algorithmic fairness, accountability, and transparency, particularly where systems automate human decision-making [5].