19PINE-AI/PreAct ? reverse-engineered prompt

Build me a Python project that lets a computer using agent get faster when it has to do the same task again.

I want the first successful run to be saved as a small reusable program, then later runs should replay that program instead of thinking through every step again. It should check the screen before each action, and if something looks different, hand control back to the agent so it can recover safely. It should also only save a run if a fresh replay really finishes the task, so broken automations do not get reused.

Make it work for mobile, desktop, and web style tasks, and include simple ways to run the benchmark experiments and compare the results with the safety check turned on and off. If it helps, look up current docs online for the benchmark environments and model setup.