Visa's agents can now shop for real in Europe.

Most of what has shipped in agentic commerce so far has been a pilot, a partnership announcement, or a protocol with no live volume behind it yet. This week Visa moved past that stage in Europe. It is now settling actual transactions where an AI agent, not a person, completed the checkout. That is a small sentence carrying a large amount of operational weight.

Pilot to production, in Europe first

Visa and a group of European banks announced they are executing live agentic-commerce transactions, with AI agents completing purchases at participating merchants on cardholders' behalf [1]. Roughly 30 European issuers are involved, and named merchants already taking agent-led purchases include lastminute.com, Frasers, BrickDepot and Cleverbridge [2]. This sits on top of Visa Intelligent Commerce, the company's existing platform for letting agents discover, initiate and complete a transaction under merchant and issuer controls [3].

Two new pieces: a score and a directory

Alongside the live volume, Visa introduced two mechanisms aimed at the trust problem underneath all this. Agent Score, built with New Generation, lets a merchant grade its own site for agentic-commerce readiness: can an agent actually navigate it, read it, and complete a task there [4]. The Agentic Directory works the other direction, listing agents and merchants Visa has verified as legitimate participants, so a merchant knows which agents it can let transact and an agent knows which storefronts are real [4]. Visa also confirmed a strategic tie-up with OpenAI, putting its network and credentialing behind agentic checkouts run inside the model itself [4].

The stablecoin coalition keeps adding names

Underneath the checkout layer, the settlement fight is still moving. Open USD, the stablecoin launched on 30 June by more than 140 firms including Visa, Stripe, Mastercard, American Express, BlackRock and Coinbase, is designed to share reserve income back to partners rather than keep it with a single issuer [5]. It goes live later in 2026 and is meant to run across multiple blockchains rather than one [6]. Separately, Visa said at the same Paris forum that it will build technology letting banks turn ordinary deposits into programmable, always-on money, so a bank can match stablecoin speed without leaving funds off balance sheet, and it has already moved stablecoin volume across VisaNet at an annualised run rate of roughly seven billion dollars as of March [4].

Read from the rails

The interesting shift is not that agents can now buy things in Europe. Pilots have shown that for months. It is that Visa built a scoring system and a directory before scaling volume, which is the correct order of operations and also the order almost nobody follows under launch pressure. A merchant readiness score is really a compatibility check dressed up as a grade, and a directory of verified agents and merchants is really an allowlist with a friendlier name. Both exist because the alternative, letting any agent transact with any site, is not a commerce problem, it is a fraud problem wearing a commerce costume.

What I would watch from here is who owns the dispute when an agent misreads a site the Agent Score called ready. A human who clicks the wrong button owns the mistake in a way a merchant and a bank both accept. An agent that misreads stock, price, or a return policy on a "certified" site creates a three-way liability question that a score does not answer by itself, it only narrows how often the question comes up. Scoring the front door was the easy 80 percent. Who eats the chargeback when the score was wrong is the 20 percent that decides whether this scales past 30 issuers.

A readiness score and a verified directory are the trust layer done properly before volume, not after. The unresolved part is who is liable when a "certified" agent still gets it wrong.