I had tried the browser screenshotting feature for agents in Cursor and found it wasn't very reliable - screenshots eat a lot of context, and the agent didn't have a good sense for when to use them. I didn't try it in this project. I bet it would work in some specific cases.