Anthropic published the code for building AI shopping agents on 2 September, and the most informative thing in it is what it refuses to do. The agent searches a catalogue, compares products, assembles a cart and then stops, handing the shopper to the retailer’s own checkout. It does not complete purchases, and Anthropic told Reuters as much on the day.

That restraint reads differently once set against the previous eighteen months of agentic commerce, in which the ambition was precisely to close the purchase inside the assistant. The company that tried hardest at that has since retreated from it, and the retreat was measured rather than philosophical.

What did Anthropic actually release?

Claude Commerce Agents is an Apache 2.0 repository, anthropics/commerce-agents, containing two agents. The shopping agent is customer-facing and handles search, comparison, cart building and policy questions. The merchant agent is staff-facing and works across inventory, pricing and campaigns, drafting changes rather than making them. Runnable examples cover retail, travel, telecom and ticketing, and a Claude Code plugin scaffolds an agent against a retailer’s own backend.

It is code, not a product. Anthropic has said it will not maintain the repository as a product or accept outside contributions. The blueprint ships with no payment protocol, no product catalogue and no advertising layer, and running it in production costs the retailer its own Claude usage on top of engineering, security and maintenance. Deployment documentation covers AWS Bedrock, Google Cloud Vertex AI and Microsoft Foundry.

The guardrails are structural rather than advisory. Merchant-side write actions are staged for human approval, product and price answers are constrained to the retailer’s own catalogue data, and the checkout step renders a cart rather than executing a transaction. For a company that markets itself on safety positioning, putting the limits in code rather than in prompts is the more legible version of that claim.

How good are the numbers attached to it?

Anthropic is leading with two figures: carts up to 35% larger and shoppers around 60% more likely to complete a purchase. Both need handling with care.

Angela Jiang, Anthropic’s head of product for the Claude platform, told Reuters on 2 September that the company had seen “cart size up about 30-35% for one partner”. One partner, unnamed, with no published methodology or sample size. These are company-reported pilot figures, not a benchmark across retailers, and the range narrowed to its top end in much of the coverage that followed.

The second number has a subtler problem. The same Reuters story carries a separate figure from Adobe Analytics, that AI-driven retail site visits convert at a 60% higher rate than traffic from other sources. Two different 60% figures, from two different organisations, measuring two different things, in one article. Secondary coverage has already begun blending them, and a marketer who repeats the merged version will be citing a conversion-rate finding about AI referral traffic as though it were evidence about agent-assisted checkout.

Reported launch users include Accenture, Mastercard and Visa, which says something about who is interested and nothing yet about what the agents do to revenue.

The release lands in a crowded field rather than an empty one. Shopify has already made agentic commerce a default for its merchants, and Google, Microsoft and Perplexity all run some form of in-assistant buying. What Anthropic is offering is different in kind: not a consumer gateway that retailers plug into, but the code for a retailer to run its own agent on its own property. The strategic logic is visible enough. Claude has far fewer consumer users than ChatGPT or Gemini, so reaching shoppers through merchants’ own storefronts routes around a distribution problem rather than competing with it directly.

Why does it stop at the cart?

Because the fuller version was tried and the numbers were poor. OpenAI launched Instant Checkout inside ChatGPT in September 2025, letting shoppers buy without leaving the conversation. Walmart joined from November with roughly 200,000 products. At Morgan Stanley’s technology conference on 4 March 2026, Walmart’s Daniel Danker said in-chat purchases converted at one-third the rate of shoppers who clicked through to Walmart’s own site, and called the experience unsatisfying. OpenAI phased Instant Checkout out later that month, and Walmart moved to embedding its own assistant, Sparky, inside ChatGPT and Gemini instead.

The failure was mostly infrastructural rather than conceptual. Product data went stale, availability and delivery estimates were unreliable, and prices shown in the conversation did not always match the ones at the end of it. The settled pattern for 2026 is discovery inside the assistant and purchase on the retailer’s own property, which is exactly the shape Anthropic has now shipped as a reference implementation.

For a B2B marketing team, the useful transfer is not about retail checkout at all. It is that these agents answer from structured catalogue data, and a code review of the blueprint found its deterministic filters miss when catalogue attributes are unstructured or units are mixed. An agent cannot correctly filter a product it cannot parse. The same logic applies to a software company’s pricing tables, feature matrices and specification pages: whatever an assistant is asked to compare, it compares using the fields it can read, and inconsistent units or prose-buried specifications quietly remove a product from consideration rather than ranking it lower.

Anthropic says it will not maintain the repository as a product or accept outside contributions.