A reference build, not a product
Anthropic published a reference blueprint on 2 September for building shopping and merchant agents on Claude. The repository carries an Apache 2.0 licence and contains two agents and four worked verticals: retail, travel, telecom and entertainment.
The shopping agent faces the customer — searching a catalogue, comparing products, building a cart, answering order and policy questions. The merchant agent faces the shop — listings, inventory, pricing, promotions and performance. Anthropic describes the package as “the harnesses, patterns, and guardrails an engineering team needs to get a commerce agent running in days,” according to Digital Commerce 360.
Agents are defined once and run across three runtimes — the Messages API, the Agent SDK and Managed Agents, the last still in beta — and the same code deploys to Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry or the Claude API directly. A Claude Code plugin scaffolds new agents.
The safety design is in the defaults
The interesting engineering decision is that all writes are staged until a human approves them. No transaction executes on its own. The repository also ships fencing, provenance gates, memory validation and approval workflows as part of the reference implementation rather than as an appendix, and the example MCP servers bind to loopback with no authentication, so a careless deployment fails closed rather than being exposed.
That is a defensible default for agents pointed at inventory and pricing, where a mistaken write is not a bad answer but a wrong price on a live storefront.
What is deliberately missing
The blueprint carries no payment protocol, no checkout and no ad layer. Anthropic has left all three to retailers and to partners, which the reporting names as Shopify, Visa, Mastercard and Accenture. Shopify and Priceline are already running live agents on the technology, and Priceline rebuilt its assistant Penny on Anthropic’s models.
Leaving out payments is the strategic choice. It keeps Anthropic out of a regulated business it would have to build compliance for in every market, and it keeps the card networks as partners rather than as parties being disintermediated. It also means the hardest unsolved problem in agentic commerce — how a merchant proves an agent had authority to buy on a shopper’s behalf — belongs to somebody else.
The numbers, and whose they are
Anthropic says retailers running shopping agents built with Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase. Those are the company’s own figures about its own product, with no independent evaluation attached and no disclosure of the baseline, the sample or the period. Treat them as a marketing claim until a retailer publishes its own numbers.
The blueprint arrives before the holiday season, which is not a coincidence: retailers freeze their stacks in October, so a reference build that lands in September is aimed at teams making a decision in the next few weeks.
What to watch is adoption outside the named partners, and whether the human-approval default survives contact with production. Staged writes are the right design and also a source of friction; the first vendor to quietly make auto-approval the default will tell you where this is heading.
