Communication infrastructure for a world of personal agents
The web is optimised for humans. If most people end up working alongside a personal model, those models are going to spend most of their time talking to each other, and natural language is a poor carrier for that. Vector representations are higher bandwidth and lower latency, and cryptography built for them becomes valuable in exactly that world.
The open question. If agents talk to each other, why are they doing it in English?
What would kill it
The thesis assumes almost everyone becomes agent-native, and that single assumption carries the whole argument. The buildable slice of it is the handover problem above.