Maybe it’s just my lack of imagination, but I don’t see a use case for giving agents money when they can barely be trusted to not break out of a sandbox and cause hallucinatory chaos.

Let’s say for a minute this is a failure of my imagination. Does each sub agent really need to be able to independently transact?

I could see a future in which ad-funded web sites charge agents some tiny amount per request instead of outright blocking non-human visitors.

Besides, you're already effectively giving agents money (tokens aren't free, and agents decide when to stop consuming new ones), this is just more explicit :)

Why not have sub agents talk to a payment service? Sub agents should ultimately be associated with a human or business entity, the human or business entity could provide a single payment service for agents.