Claude Commerce Agents: the agent got easy to build. The catalog decides what it can enforce.

On 2 September 2026, Anthropic published Claude Commerce Agents: an open blueprint for a shopping agent in the storefront and a merchant agent for the back office. It includes reference implementations for retail, travel, telecom and ticketing, plus a Claude Code plugin to adapt the pattern to a company's own systems. Shopify published its implementation on the same day, connecting a storefront agent to a live Shopify store through the Universal Commerce Protocol.

This is why the conversation has changed. The agent layer is becoming practical infrastructure for discovery, conversational shopping and purchase flows. Anthropic reports that retailers already running shopping agents on Claude have seen larger carts and higher completion rates. Those are Anthropic's figures, not ours. The useful signal is that a retailer can now start from a working architecture rather than a blank page. That is the operational side of agentic commerce.

We cloned the repository, ran the retail example and read the code behind the shopping flow. It makes one thing very clear: the agent can only apply the product conditions that the catalog gives it in a usable form.

The blueprint ships the agent, not the catalog

Claude Commerce Agents supplies the agent loop, the skills, the tool contracts, memory, presentation rules, safety gates and evaluation harnesses for both sides of commerce. A team can run that pattern on the Messages API, the Claude Agent SDK or Claude Managed Agents. The shopping agent covers discovery, purchase research, planning, customer care and personalisation. The merchant agent covers performance, listings, inventory, pricing and campaigns.

The retailer's catalog is outside that blueprint. StorefrontBackend is an interface that each company connects to its own product, cart, order and policy systems. The retail example uses 87 fictional ACME products across 11 categories. The AI shopping agent sees only the product record that this connection returns.

How the Claude Commerce Agents shopping agent searches your catalog

The shopping agent does not inspect the catalog on its own. It calls search_products with a query, filters and a limit of eight results. The important part is how the blueprint tells it to construct that call.

The search skill is direct: use the catalog's vocabulary for the query, and put each condition the shopper explicitly states into a filter. A request for a bed frame with nine inches of clearance and a 700 lb weight capacity becomes a search for "platform bed frame", a price ceiling and two product-attribute filters. The query finds candidates. The filters decide whether candidates qualify.

The retail example shows why this distinction matters. It ranks text with the title weighted highest, then brand, category, attributes and description. It keeps results within half of the top score. Only afterwards does it apply filters. Those filters inspect attributes, titles and option values. They do not inspect the specification sheet or the description.

If a filter would return nothing, the demo backend removes it and sends back the ranked list. The agent is told that those results may be related items, not the requested product. A shopper condition remains enforceable only when the catalog holds it as a field.

A production backend may use Elasticsearch, Algolia, Factfinder, a vector index or another search service. The ranking will differ. The requirement does not: product conditions need fields, text ranks candidates, and the agent can only work with the record search returns.

Three things this asks of your product data

A stated condition has to be a field, in the words the agent will use

In Anthropic's demo catalog, 63 facts sit in the spec sheet of 29 products and nowhere else: under-bed clearance, weight capacity, tent floor area and whether a suitcase has a built-in lock. A shopper can read those facts on a product page. The filter cannot use them, so a request that depends on one may return a shortlist of look-alikes instead of a qualified product.

We moved those facts into product attributes with a 48-line script. The merchant agent's catalog-listings skill follows the same discipline: it can fill a missing attribute from a value already in the record, and it stops when the record does not contain the fact. That is a small example of the difference between writing product copy and creating a product fact an agent can use.

The value has to be comparable

A field alone is not enough. The demo filter compares strings. weight appears on 14 products as 254 g, 2.1 kg, 35 lb, 2 lb and 2 x 9 oz. size can mean a bed size, a rug dimension, a storage cube, a bottle or a grocery pack. The bed frame carries weight_capacity: 700 lb, and a filter for "700 lbs" does not bind, so the backend drops it and returns a ranked list.

Structured product data needs both a field and a comparable value. That means one unit, one format and one allowed vocabulary for each category. It is the difference between information a person can interpret and information a system can apply.

What lives in the description is invisible until someone asks for it

A search result gives the agent a title, price, attributes and short description. The specification sheet and long description arrive later, one product at a time, through get_product_details. In the small retail demo, the right product is usually among the eight results and a careful agent can check it before answering.

An enterprise catalog changes the stakes. A product whose deciding fact lives only in prose has to win a text-ranking contest before the agent can inspect it. A fact that the filter cannot read never becomes a condition on the shortlist.

Making it a field is the beginning

On the demo catalog, promoting a fact into a field is a script. In an enterprise retail catalog, the same move raises the questions that decide whether the field can be trusted: which attributes a category needs, where each value comes from, which variant it belongs to, and which version may be used in each market and channel.

Some facts are accurate in one context and unusable in another. The demo sunscreen has water_resistant: 80 min, a US-style claim. Its description also says "reef-safer formula" without an evidence field behind it. In Europe, environmental and product claims face tighter consumer-protection rules from 27 September 2026. The merchant-agent guardrails already reflect the principle: some compliance and tax information cannot be changed in bulk by an agent.

That is product data governance. A retailer defines the attributes each category requires, connects every value to a source, keeps it current, normalises the units and allowed values, then applies the correct version by variant, market and channel. Product data quality is an operating discipline, not a cleanup exercise. It is what makes product data for AI agents reliable.

An agent ready catalog is one where every fact a shopper may state as a condition exists as a field, holds a comparable value and carries the rules for where it can be used. Creating the field solves the interface problem. Keeping the value trustworthy across the catalog is the enterprise work.

CommerceClarity works alongside the PIM, ERP, DAM and supplier systems a retailer already runs. Catalog Foundation turns scattered product information into usable records. Context connects those records to company rules, systems, markets and category requirements. Compliance and Quality applies those rules before a value reaches a channel.

Where to start

Start with the product questions the agent needs to answer. Take them from on-site search, customer care, merchandising and the sales floor.

For each question, ask whether the catalog can identify the right product rather than a plausible shortlist. Check whether the deciding facts are fields with comparable values. Then check whether those facts are valid for the right variant, market, channel and category rules.

Fix one category, validate it against a golden set, then extend the work category by category and channel by channel. The agent is becoming easier to build. The catalog work decides what it can enforce.

If you want to run that check on your own catalog, book a pilot.