Claude for Commerce gives every retailer a shopper agent. Ask whose shopper it serves.

TL;DR

On 2 September Anthropic open-sourced two reference commerce agents: a shopping agent a retailer embeds in its own app or website, and a merchant agent for the back office. AI agent shopping now has a free blueprint, and its system prompt opens by naming the brand it works for, tells the agent never to point a customer to another named retailer, and stores what it learns about you in the store's systems. That is the store's agent, well built and honestly described, and it is not the shopper's. Anthropic's own engineering post says the shopper's agent is next: "some of the traffic to your storefront will come from agents that shop on behalf of users."

What Anthropic released on 2 September

Start with what anyone can check. Anthropic published a blueprint for building commerce agents and a reference repository under Apache-2.0. The README describes the two agents in it as "a shopping agent a business embeds in its app for customers, and a merchant agent its staff use to run the back office." Each runs on three paths (the Messages API, the Agent SDK, Managed Agents), across four verticals with eight demo apps, and a Claude Code plugin scaffolds a new one against your own systems. Shopify and Priceline are named as customers already running agents on Claude. Accenture, Mastercard and Visa are named as partners who will deploy the blueprints for their clients, as Digital Commerce 360 reported the same day.

Retailgentic covered the launch the same afternoon and gave the first agent the right name: a retailer-hosted shopper agent. Keep the adjective. It is the whole argument.

Whose agent is it? Read the system prompt

Nobody in the launch coverage quoted the code, so here is the first line of the shopping agent's system prompt, from shopping-agent/core/shopping_agent/prompt.py in the repository:

You are {config.assistant_name} for {config.brand_name}, talking with a customer inside the store's app or website while they shop.

The default assistant_name is "the shopping assistant". The brand fills in the rest. Three more rules follow in the same file, and the README adds a fourth. Each one is exactly right for a store's agent.

• Stay inside. The scope line reads "Stay within {scope} for {config.brand_name}", where scope is shopping, planning, orders and the store's terms.
• Never name elsewhere. "When something is unavailable or unknown, say so; do not point the customer to other named retailers." And: "Do not name a real business, venue, or brand outside this catalog".
• A gap is a fact about the store. The customer "hears about a catalog gap as a fact about what the store carries."
• Checkout is the store's. "Nothing places an order, charges a card, or changes a live listing: checkout renders the cart for the host to complete."

Be fair about the intent, because it is good. The same prompt says "Recommend what fits the customer's stated needs and budget and name the trade-offs. You are not there to promote." The research skill tells the agent to read the store's own buying guide before it recommends anything. The safety document lists what the harness enforces so that, in Anthropic's words, "No model tool call moves money or changes the business." This is careful engineering and it does not pretend to be anything it is not.

The 4GPTs read: this is the store's agent. Named for the store, scoped to the store's catalog, forbidden from pointing outside it, handing off to the store's checkout. All of that is correct for the job it has. The only mistake available is in the name, and the name is where the launch coverage went: a shopper agent for every retailer. A shopper agent the retailer hosts is the retailer's agent wearing the shopper's clothes. It serves you inside the store. It cannot serve you when the right answer is the door.

Where your shoe size goes

The memory design makes the same point from another angle. Anthropic's engineering post is explicit: "Memory belongs in your systems, not in the model." A memory is "a key (such as shoe_size, default_store, preferred_report_cadence), a short value, a category, and the session it came from." It is stored under the customer id the store's own login resolved, with a retention period the store sets, and "Memory should be a per-deployment switch."

That is what retailer-hosted costs the shopper. The requirement, the sizes, the budget, the thing you will not compromise on: those are yours, and the natural place to hold them once is an agent you run.

Two numbers with no base

The announcement says retailers running shopping agents on Claude "have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase." Digital Commerce 360 passed the 60 along. Neither figure comes with a retailer count, a period or a comparison group, and "up to" is a ceiling from an unstated set.

It is the same shape as the 620 percent assistant number from a fortnight ago. A shopper who opens an on-site assistant has already decided to buy something, so a conversion multiple against everyone else measures who clicked before it measures what the software did. None of that makes the figures false. It makes them a launch metric.

What Retailgentic gets right

Two things, and both matter more than the launch itself.

The name, already noted. And the foundation. Retailgentic reads the product cards Claude began showing in June as evidence of "a whole ontology" underneath: a canonical catalog, hero SKUs, a taxonomy, feeds and crawl, in its words "essential foundational building blocks for building anything in our world." That is right. An agent on either side of a sale is only as good as the product data it can read, which is why agents buy what they can read and why the 4GPTs mechanism runs on catalog and price data fresh enough to act on. Whoever does the ontology work well is building for both sides at once, whatever the agent on top is called.

Two jobs, two agents

The most useful paragraph in the whole release is the last one in the engineering post, and the coverage skipped it.

Read it twice. "Your own agent" and "agents that shop on behalf of users" are two different things, and the vendor says so. The store's agent answers for the catalog. The shopper's agent holds the requirement and shops the market against it. One is now a free blueprint. The other is named as the next arrival.

The 4GPTs read: this release moves the house bet by one step, on one side. The supply side of the agentic agora is arriving owned by the merchant, and cheaper than anyone priced it: Retailgentic's estimate is "probably 200 companies building retailer shopper agents", and Anthropic just published the code. The demand side is untouched by it, and the shopper's agent still holds the one thing a hosted agent cannot: the walk-away. What sits between the two is the part nobody shipped this week, a connection that lets a person's agent put a stated requirement to a store's agent and get a real offer back. Both sides of the sale are agents. 4GPTs is the layer that connects them, and this release makes one side a commodity while telling every retailer to prepare for the other. Today 225 of the thousand largest US retailers expose an endpoint an outside agent can detect. The clock on the other 775 just got shorter.

What a merchant does on Monday

1. Deploy it as the store's agent, and say so. The blueprint is good. Name it what it is in your own roadmap, so nobody on your team confuses it with the agent your customers will bring.
2. Treat the outside agent as your next customer. Anthropic told you it is coming. A machine-readable catalog and an endpoint it can transact through are the work, and the provenance and approval rules you deploy for your own agent are, in Anthropic's words, what let you open your tools to it safely.
3. Keep the memory boundary honest. Show people what your agent holds about them, let them delete it, set the retention. The blueprint asks for all three. Do them before launch, not after.
4. Publish the base with the ratio. If your assistant grows carts, say by how much and from what. A number with no denominator invites the question this article is.

FAQ