Tag
PreAct compiles successful agent runs into small state-machine programs, enabling 8.5-13x faster replay on repeated tasks without per-step language model calls, with runtime screen checks to ensure correctness.