Comment by angry_octet

I think you're right, you can get agents to do what we do -- learn how a website works. Then expose that model as a simple API. There will still be some vision tasks for navigation but they will be just vision tasks, no thinking required.