Anthropic has published a blueprint that lets any retailer build an AI shopping agent on Claude. The release contains two reference agents: a customer-facing shopping agent that searches a catalogue, compares products and assembles a basket, and a merchant agent that sits on the other side of the store handling inventory, pricing and marketing. It ships with runnable examples across retail, travel, telecom and entertainment, a safety harness and a set of approval gates.

What it does not contain is just as telling. As PYMNTS noted, there is no payment protocol, no checkout and no advertising layer. The agent builds the cart and hands the shopper to the retailer’s own checkout. Nothing routes through Anthropic.

This marks another step towards AI becoming embedded in the transaction journey. After OpenAI’s push into agentic commerce, Anthropic is now putting the technology directly into retailers’ hands, allowing them to build agents that can recommend products, assemble baskets and support customers through to the point of purchase.

I would be cautious, though, about interpreting that as a future in which every retailer needs to own the shopping agent.

Consumers have already decided where they want the agent to sit

Recent PSE Consulting research across 4,250 consumers in the UK, US, France and Germany found that 74% prefer an independent AI assistant for shopping, split between a universal tool that works across multiple platforms and providers and a specialist assistant focused on a category such as travel, finance or healthcare. Only 10% want AI embedded within a single retail, travel or delivery platform.

The pattern underneath that number is consistent. Consumers want an agnostic agent at the top of the funnel, helping them search, compare and build a shortlist, before moving to a merchant or marketplace they trust to complete the purchase. They are willing to hand over the tedious part of shopping. They are noticeably less willing to hand over the part that carries financial and emotional weight.

Anyone who watched OpenAI’s commerce strategy over the past year will recognise the shape of this. Instant Checkout launched with considerable fanfare and was then scaled back in March in favour of checkout experiences controlled by merchants. Discovery in the assistant, purchase on the merchant’s own property, turned out to be the durable pattern. Anthropic’s blueprint appears to have been designed with that lesson already absorbed, which is why it stops at the basket.

Two agent layers, not one

That said, I can see a world where independent and on-platform agents coexist comfortably.

Retailers and marketplaces have a strong incentive to build capable agents, because they can use proprietary data, existing customer relationships and fulfilment infrastructure to personalise the final stages of the journey in ways an agnostic assistant cannot. Order history, loyalty status, stock position in a specific store, returns behaviour and delivery options all sit inside the merchant’s own systems. Marketplaces also have the scale to invest heavily in agent development and then deploy those capabilities across very large customer bases.

Anthropic has lowered the cost of prototyping that capability. It has not removed the expensive parts, which are catalogue quality, systems integration, safety engineering and ongoing operations.

That creates a more difficult environment for smaller merchants. Building a sophisticated proprietary agent requires data, technology and investment that many smaller businesses will struggle to match. Independent agents could therefore become just as important to merchant visibility as retailer-owned agents are to conversion. For a small business, being present and well described in the shortlist an independent assistant produces may matter far more than owning an assistant of its own.

Trust is the constraint on all of this

The wider context matters here, because agentic commerce is arriving in the middle of a serious industry conversation about whether agents can be controlled at all.

The Hugging Face incident, in which OpenAI models circumvented controls designed to isolate them from the internet and compromised production infrastructure, was the clearest demonstration yet of what happens when autonomous systems pursue an objective without adequate containment. The independent investigation by METR made uncomfortable reading for anyone planning to give agents access to live systems. Days ago, Anthropic’s own chief executive published an essay arguing that the industry should slow down and submit to third-party evaluation and eventually federal regulation, a position quickly echoed by his counterparts at OpenAI, Google DeepMind and xAI.

None of that is a reason for merchants to sit out agentic commerce, it is a reason to be precise about what an agent is allowed to do on your systems and on your customers’ behalf. The payments industry has spent decades building exactly the machinery this problem needs: mandates, authorisation, spend controls, dispute rights, auditability and clear liability when something goes wrong. Those questions are becoming more pressing, not less, as agents move from recommending products to acting on instructions.

It is worth noting that Anthropic’s blueprint reflects some of this thinking, in that merchant-side changes are staged for human approval rather than executed automatically. The design assumption is that a person stays in the loop where money and live listings are involved.

What this means for merchants

Consumers stand to gain greater choice and more useful assistance across the shopping journey. Merchants gain new ways to convert and retain customers, but they will need to compete on two fronts at once: for inclusion in independent AI shortlists, and for engagement once a shopper reaches their own platform.

Those are different disciplines requiring different investments. The merchants who do well over the next two years will be the ones who understand that the discovery layer and the execution layer are separate battlegrounds, and who resist the temptation to assume that owning an agent is the same thing as owning the customer.