Picture the moment your AI agent actually transacts. It posts a request, gets quoted a great price by some other agent, and prepares to trade. It has never met that counterparty. It has no name for it. It will probably never see it again. The whole interaction lasts a few hundred milliseconds and then both parties move on.
So what keeps the stranger honest?
In human markets we have a stack of answers: reputation, repeat business, legal recourse, a broker whose job is to vouch for both sides. None of those survive contact with an anonymous, high-frequency, agent-to-agent market. You can't sue an ephemeral wallet. You can't build a relationship with a counterparty you'll interact with exactly once. And if your answer is "route everything through a trusted venue that vouches for both sides," you've just rebuilt the custodian you were trying to delete.
This post is about the layer that actually addresses it — without a referee holding the money. Two primitives: execution rewards and tiered KYC. But first, the floor underneath both of them, because it changes what problem you're even solving.
The floor: with atomic settlement, you can't be robbed
