Abstract
Existing Graphical User Interface (GUI) agents operate through step-by-step calls to vision language models--taking a screenshot, reasoning about the next action, executing it, then repeating on the new page--resulting in high costs and latency that scale with the number of reasoning steps, and limited accuracy due to no persistent memory of previously visited pages. We propose ActionEngine, a training-free framework that transitions from reactive execution to programmatic planning through a novel two-agent architecture: a Crawling Agent that constructs an updatable state-machine memory of the GUIs through offline exploration, and an Execution Agent that leverages this memory to synthesize complete, executable Python programs for online task execution. To ensure robustness against evolving interfaces, execution failures trigger a vision-based re-grounding fallback that repairs the failed action and updates the memory. This design drastically improves both efficiency and accuracy: