browser-use/jev-ultrafast ? reverse-engineered prompt

Build me a fast browser automation agent that can take one plain English goal and carry it out in Chrome, then show me what it’s doing in a local web dashboard.

I want it to read the page in a structured way, figure out the best next action like click, type, select, scroll, wait, or finish, and only generate text when it actually needs to type something. It should work well for tasks like searching flights or opening a Wikipedia article, and it should verify that the goal was really completed instead of just guessing.

Please make it runnable locally with a simple setup, using environment variables for the model keys, and include a small demo and a few examples. I’d also like a clean Python library interface so I can use it from my own script.

If you need to look up current docs for the browser connection or model setup, go ahead and check online first.