Anthropic just open-sourced shopping and merchant agents. Here is what they still do not give you

What actually shipped

The repository is public at github.com/anthropics/commerce-agents, licensed Apache 2.0, free to fork and modify. It contains two working agents rather than one. A shopping agent sits inside a retailer's own app or site and handles the customer-facing loop: search the catalogue, compare options, build a cart. A merchant agent works the other side of the counter, helping store operators with inventory, pricing and marketing decisions. Anthropic describes the package as the harnesses, patterns and guardrails an engineering team needs to get a commerce agent running in days rather than months.

The reference examples span four verticals, retail, travel, telecom and entertainment, each with its own working example and a companion staff-facing portal. The release also ships a plugin for Claude Code, Anthropic's developer tool, so a team can scaffold a new vertical from the existing pattern instead of starting from the protocol spec.

What Anthropic deliberately left out

The blueprint ships with no payment protocol, no product catalogue and no advertising layer. Those three pieces stay with the retailer's own systems and payment partners, so no transaction routes through Anthropic itself. That is a narrower scope than it might sound. It means the hardest, most regulated parts of a commerce agent, moving money and knowing what is actually in stock, are explicitly somebody else's problem: yours, or your existing payment and inventory vendors'.

Adoption so far reads the same way. Shopify and Priceline are both already running customer-facing agents on Claude; Priceline rebuilt the latest version of its assistant, Penny, on Anthropic's models. Those are production deployments layered on top of platforms that already owned payments and catalogue. The blueprint is the connective tissue, not the whole animal.

The shopping agent: the same gap, dressed differently

We wrote about this exact problem two weeks ago when Shopify's WebMCP went live on every Liquid storefront. The shape repeats here, for a different reason. WebMCP is free and automatic, so the gap is that every store got the same verbs on the same day. Claude Commerce Agents is a blueprint a team has to actively build, so the gap is different in kind: even after the engineering work is done, the agent Anthropic's own reference implements can search, compare and add to cart. Nothing in the harness tells it why one of four comparable products deserves the sale over the others.

That is not a criticism of the blueprint. It is honest about its own scope in a way a lot of agent-framework marketing is not. But it means a retailer who ships Claude Commerce Agents this quarter and stops there has built a faster, better-engineered version of the same evidence gap every other agent surface has: complete Product schema, an AggregateRating backed by real reviews, and rights-cleared, product-tagged customer photos and video the agent can quote when it explains a recommendation. None of that arrives with the repo, and it was never going to. It is catalogue-and-content work, not model work.

The merchant agent: the half nobody is writing about

Most of the coverage of this release, understandably, is about the shopping agent. It is the customer-facing half and it is the one that maps onto WebMCP and ChatGPT Shopping and everything else in the agentic-commerce conversation this year. The merchant agent is the quieter half, and it raises a question the shopping-agent framing does not.

A merchant agent making inventory, pricing and marketing calls is only as good as the signal it is reasoning over. Ask it which SKUs to feature in a promotion and it needs to know, with some confidence, which products actually have strong customer sentiment behind them, not just which ones are overstocked. Ask it to draft marketing copy and it needs source material it is actually allowed to use: which customer photos and reviews carry a usage right that covers a paid placement, and which ones are display-only. Feed a merchant agent an untagged, unrights-cleared content library and the most likely failure mode is not a bad answer. It is a confident one, built on the wrong evidence, that a human approves because the agent sounded certain.

What to do about it this week

1. 1Read the repo before the press coverage. github.com/anthropics/commerce-agents is public and the READMEs are short. Thirty minutes tells you more than any third-party writeup, including this one.
2. 2If you are already building a commerce agent with a router and subagents, read the architecture note above properly. One agent loop plus skills is a deliberate choice in Anthropic's own reference, worth weighing against whatever pattern your team defaulted to.
3. 3Do not let "no payment protocol, no catalogue" become "not our problem yet." Those are explicitly scoped out because they are supposed to be already solved, by your existing stack. If they are not, that is the actual blocker, not the agent framework.
4. 4Audit the evidence layer before you build the shopping agent's product-selection logic. Product schema completeness, attributable reviews, and product-tagged UGC. Our free AI Visibility Check grades it on any URL.
5. 5If a merchant agent is on your 2026 roadmap, get the rights and tagging layer in place first. An agent reasoning over your content for pricing or marketing decisions needs the same consent and provenance record a human marketer would want before publishing the same material.