Three cycles. Rail. Electrification. The internet. Survivors weren't the fastest adopters. They separated structural change from noise before anyone else thought the distinction mattered. AI is not different in kind. Only in speed.

An agent filling a cart makes one decision: accept the posted price or leave. There's no counter, and no walk-away that costs the seller anything. The price was set by one side, and who captures the surplus was settled before the agent was switched on.

A trading mandate has a size limit and a named person the exceptions go to. An agent with payment credentials has a rate limit. When a payment falls outside the owner's stated limits, which human is it meant to come back to, and what happens to it while it waits?

Every agent commerce demonstration I've seen shows a purchase completing. The useful one would show it failing halfway: money gone, nothing delivered, and someone deciding who carries the loss. The reversal path is the product. Nobody is demonstrating it.

Escrow, and a party who can reverse the settlement under rules both sides agreed to beforehand. Markets have cleared at speed on that arrangement for a long time without either counterparty being trusted.

An agent that fills a cart and pays is doing order entry. The seller set the price. The agent only chose yes or no. Price discovery needs an offer, a counter, and a buyer who can credibly walk away. Without the walk-away, a smarter buyer just accepts faster.

Agent payment credentials from the card networks: by the end of 2026 I expect a per-transaction cap and a human approval step above it, on by default. I'm wrong if the default limit is still whatever happens to be sitting in the funding account.

Delivery versus payment exists because someone once handed over the shares and never got paid. The agent demos all show the negotiation. None of them show what happens when the goods move and the money doesn't, and which side is left holding a claim on the other.

An agent buys from another agent. A week later the charge is disputed, and the dispute lands on whoever's name is on the card. Nobody has yet said which side has to prove the instruction was authorised, or what that proof looks like when both parties are software.

A clearing house trusts neither side of a trade. It holds the money and can undo the deal. That is the whole arrangement. Agent commerce ends up in the same place, and the work gets done by the venue, not by anything anyone does to the agents themselves.

Most of what's shipped so far is an agent filling a cart at a price the seller already set. That's order entry. A price gets discovered when a counter-offer is possible and walking away costs the other side something. Neither thing exists in a checkout flow.

Anyone handed authority over someone else's money gets a cap and a name to ring above it. An agent with a card and no ceiling has neither. The loss still lands on the owner, who was never asked to put a price on it, because nobody wrote the limit down.

Goods released, payment later reversed. That's the shape I expect the first agent-payment failure large enough to be written up to take. Scoreable by the end of 2026. If the first one turns out instead to be a stolen key or a model talked into paying, I was wrong.

A clearing house trusts neither side of the trade. It holds the money and it can unwind the trade when one side fails. Agent commerce keeps getting discussed as a problem of making the agents behave. Escrow and rules agreed in advance do work that behaviour can't.

A payment gets authorised by software acting for a buyer, the goods never arrive, and the chargeback has to land on someone. The card rules were written around a cardholder who can say he didn't authorise it. Who says that now, and who is believed?

A cart with the price already in it is order entry. Price discovery needs an offer answered by a counter, with both sides able to walk away. Most agent commerce so far hands the agent the last click, and whoever set the number keeps the surplus.