Anthropic ships a working shopping agent, and twenty-one banks commit to a dollar stablecoin for 2027

Anthropic published blueprints yesterday for building shopping and merchant agents on Claude, aimed at retailers going into the holiday season [1][2]. Not a product. Working reference implementations, across retail, travel, telecom and ticketing, shipped as something an engineering team clones [2].

The shopping agent searches a product catalogue, assembles multi-item carts, tailors recommendations to a customer, shows comparisons inside the conversation, hands the finished cart to checkout, and answers questions about orders, returns, exchanges and refund policy [2]. The merchant agent answers questions about sales performance, watches inventory, flags problems, recommends pricing and promotions off sales history, and drafts marketing campaigns [2]. Either one deploys through the Claude API, Amazon Bedrock, Microsoft Foundry or Google Cloud Vertex AI, and there is a Claude Code plugin [2].

Angela Jiang, Anthropic’s head of product for the Claude platform, told Reuters that cart size is up about 30 to 35 per cent for one partner, and that customers are 60 per cent more likely to complete a purchase [1]. One partner, unnamed, no baseline published, no window given. Adobe Analytics separately has AI-driven visits to retail sites converting at 60 per cent above other traffic [1], which is a different measurement of a different population, and the two do not stack.

The merchant agent still asks permission

Campaign drafting requires human approval before anything goes live [2]. It is the only place in either agent where the blueprint stops and hands the decision back to a person.

That split turns up everywhere in this work. Read paths get automated first and write paths get a human in front of them, because a wrong answer about last week’s sales costs somebody a second look, and a wrong discount emailed to a list costs margin nobody claws back. Anthropic has drawn the line inside the reference implementation, so it does not get drawn instead by whoever wires up the tools on a Thursday afternoon.

What the box does not carry is the money. The shopping agent hands a completed cart to checkout [2]. Checkout is where the agent’s identity, the consumer’s authorisation record and the eventual dispute all have to live, and none of the three ships in this release.

Twenty-one banks and a company nobody has named

Twenty-one banks and asset managers committed on 1 September 2026 to form a company in the second half of this year, subject to closing conditions, and to bring a US dollar stablecoin to market in the first half of 2027 [3][4].

The North American names are Bank of America, Capital One, Citi, Fidelity Investments, Goldman Sachs, PNC Financial Services, Scotiabank, TD Bank Group, Wells Fargo and WisdomTree. The European ones are Banco Santander, BBVA, Commerzbank, Crédit Agricole, Deutsche Bank, Lloyds Banking Group, Rabobank and UBS. MUFG Bank, Sirius International Holding and Standard Bank cover East Asia, the Middle East and Africa [4].

The token is meant to serve wholesale, institutional and retail payments alongside digital asset settlement, and the venture says it intends to meet the applicable GENIUS Act and MiCA requirements before it operates globally [3][4]. A euro-denominated token is named as the next priority, with other G7 currencies behind it [3][4].

There is no company name, no token name, no blockchain and no custodian [4]. That is the state of the announcement rather than a complaint about it, and it is why the 2027 date belongs in a diary and not in a roadmap. Getting twenty-one institutions to agree on who holds the reserves and who eats a redemption run at three in the morning is harder than anything in the technology, and it is precisely the part they have not published.

Mastercard has named twenty-two of the gaps

Mastercard’s Start Path took in 22 companies on 2 September 2026 for the first cohort of its Agentic Commerce and Services track, which it opened in January [5][6].

Read the list as a map of what is still missing between an agent and a settled payment. Skyfire gives agents verified identity and payment credentials for use across the open internet. Merge connects agent tools and AI models to business systems. Firmly lets merchants sell through AI agents without additional engineering. ReFiBuy reshapes product catalogues so shopping agents can find what is in them. Nevermined verifies, accepts and settles agent-driven purchases on existing payment rails [5]. The rest of the cohort is Crossmint, SolvaPay, Tunic Pay, t54, PayOS, Simetrik, Wizard, Grovity, 10Clouds, Sapiom, Satsuma, Glassbox Labs, Masraff, MediConCen, Luca AI, Thyris.AI and Testudo [5].

Start Path is an accelerator, so the signal here is Mastercard’s shopping list rather than any individual company’s prospects. Agent identity, catalogue formatting for machine buyers and settlement on rails that already exist are three things a card network has decided not to build in-house this year.

The blueprint is the part that is ready. Anthropic can hand a retailer a shopping agent that builds a working cart this week, and that cart arrives at a checkout where the agent still carries no verified identity and the customer’s authorisation still lives as a variable inside somebody’s application. The dollar it would rather spend belongs to a company that has not been formed yet, and whose token is not due in the market until the first half of 2027.