When AI generates code, I first instruct the model to find, fix, and verify any issues. After that, I start the server and test whether it actually works from the user’s perspective. What I’m looking for is a workflow where issues are received, fixed, tested, and deployed—but it seems that current AI agents aren’t very good at performing browser tests from the user’s perspective. I’ve tried using the built-in browsers in Codex and Cursor, but they often only checked whether the page loaded. In t...