why would you make an agent click around a web browser like a human when it could self discover the api and call it directly?

self discovery via primitives is what works well today. I never discouraged that, only discouraged MCP sensationalism.

However, an agent that can see the screen and immediately click through whatever desired UI modality is immensely more efficient than swimming through protocols.

There is at least one frontier lab who has prepared enough foresight that agents running on VDI infrastructure is a major coming wave.