The useful version of this for me is less "agent can browse a website" and more "agent can keep the browser state in the same work loop as the repo." A lot of real tasks are half local files, half logged-in web UI, and the handoff between those two is usually where the work gets weird.