An agent can pay a Coinbase merchant today, and most bank AI still only reads and scores

Coinbase Business now accepts payments from AI agents. The agent goes through the same checkout a person would, pays over x402, and the money settles into the merchant's account as USDC [1].

There is no separate agent lane for the merchant to build. The endpoint was already there.

What actually shipped

PYMNTS carried the announcement on 11 August 2026 and FinTech Global on 12 August [1][2]. x402 is an open standard for machine-to-machine payments, and Coinbase has wired it into the checkout its merchants already run, with settlement in USDC [1][2]. USDT is now accepted across links, checkouts and invoices, converting automatically and settling as USDC [1]. Payment links are reusable, so a seller sends one link by email, text or QR code instead of minting one per customer, caps how many times it can be used, and pauses it when they want to [1][2]. There is a minimum and maximum amount, or a range the buyer chooses inside, plus a product catalogue that writes the product details once and reuses them, and the checkout collects a name, an email and a shipping address on the way through [1]. Coinbase Business serves more than 5,000 companies and the acceptance suite has handled more than 100,000 payments [1][2].

I have spent enough hours inside merchant integrations to know which line on that list I would look at first. It is the usage cap on a reusable link.

A cap is the only control in that feature set that works without the payer's cooperation. Agentic checkout so far has mostly assumed the agent behaves: a mandate it respects, a limit it honours. A link that stops working after the fifth use stops working after the fifth use, and it does not care whether the thing hitting it is a person, a cron job or a model having a bad afternoon.

FinTech Global reports Coinbase saying its fees come in under card fees and that merchants face no chargebacks [2]. The second half of that is worth sitting with, because no chargebacks runs in both directions. If the buyer is an agent acting for a company and it buys the wrong thing, there is no reason code, no representment and no acquirer standing in the middle. Someone at the buying company owns that loss, and I would be surprised if many of them have worked out yet that this is the trade.

Where the regulated side points its models

Deloitte's 2025 EMEA Model Risk Management Survey went to 136 institutions, 87 banks and 49 insurers across Europe, the Middle East and South Africa, fielded between July and September 2025 [3]. Two-thirds of them run models with AI or machine learning inside them, and that share has grown since the 2023 round [3].

What those models do is the part I keep rereading. Fraud detection, which includes anti-money laundering and know your customer, is the leading application at 58 per cent of banks and 30 per cent of insurers. Customer experience comes next, at 53 per cent of banks and 37 per cent of insurers [3][4].

Both of those are jobs where a model reads something and returns a judgement, and a rule or a human does whatever happens next. Neither one is a model that decides to spend.

The gap is in model governance

So the rails for agent-initiated payment are live at a company with more than 5,000 business customers, and the institutions holding everyone's money have their AI mostly pointed at reading and scoring.

My read is that this is a governance shape rather than a speed problem. A model that scores a transaction can be validated against what happened afterwards: you have the outcomes, you have the labels, you can measure how often it was right. A model that initiates a transaction has to be signed off before it acts, on a decision it has not made yet and against an outcome that does not exist. Model risk frameworks were written for the first kind. That is a design gap rather than a question of pace, and better rails do not close it.

Which is why near-term agentic payment volume will keep turning up where Coinbase has just put it. A merchant taking money needs a checkout endpoint. An institution sending money needs a model validation committee, and those two things move at very different speeds.

Read from the rails

Two stories today rather than three. I went across agentic commerce, the card networks and stablecoin regulation for the last 48 hours and could not find a third that was both new and checkable at a primary source. One candidate went out on the check: a widely repeated survey figure on stablecoin priorities inside financial institutions, where the study page I could reach describes a different sample from the one carrying the number. I would rather run short than print a figure I cannot stand behind.

Two things worth doing this week. If you sell anything and you have a Coinbase Business account, go and look at what a reusable link with a usage cap would replace in your current flow. It is a much smaller change than agent checkout and it is available now. If you work anywhere with a model risk committee, find out whether your framework has any language at all for a model that initiates a transaction rather than scoring one, because that is a different question from whether the model is accurate.

So which side of the gap are you on? The one where an agent can already pay you, or the one where your own agent still needs a committee?