Digital Insights: Claude Commerce Agents
Anthropic has released its Blueprint for Commerce Agents. Shopify and Shopware have already showcased prototypes. Christian Schilling explains what’s behind the architecture and what this means for companies with extensive product catalogs.
On September 2, 2026, Anthropic released its Blueprint for Commerce Agents. Less than 48 hours later, Shopify’s implementation was already publicly available on GitHub.
Claude Commerce Agents is an open-source reference architecture (a blueprint) from Anthropic that enables companies to develop AI agents for online retail and e-commerce.
In its announcement, Anthropic cites up to 35 percent larger shopping carts and a 60 percent higher likelihood of a purchase being completed. This hasn’t been proven, but it’s promising. It’s definitely worth a closer look for merchants. Not because an AI agent will be ordering your products tomorrow, but because this demonstrates what a technology that truly deserves trust looks like—and what that means for your own preparations.
Not a new model, but a blueprint
Important to understand: Anthropic has not released a new AI model. The company itself describes the release as a development framework. The repository contains ready-made reference implementations for a shopping agent and a merchant agent, along with security mechanisms, testing procedures, and a plugin for Claude Code. Four industries are covered: retail, travel, telecommunications, and ticketing. The code runs via the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. Visa, Mastercard, and Accenture are on board as ecosystem partners.
Two agents, two tasks, one clear boundary
The shopping agent resides in the online store or app and helps users search, compare, and fill a shopping cart. The merchant agent operates on the back end, handling inventory levels, pricing suggestions, and campaign drafts. The two agents are deliberately kept separate.
This aligns with what we’ve learned from our own projects with clients offering a wide range of products. Customer communication and internal management require different approval processes, different risk profiles, and, in most cases, different in-house personnel responsible for each.
The real breakthrough: Control lies in the code, not in the prompt
The most interesting architectural decision does not concern the agent’s language, but rather its limits of action. A Merchant Agent may suggest a price, but it may not approve it on its own. Every write operation first runs as a draft—visible as a “before-and-after” comparison—and takes effect only after human approval. Products and transactions are accessed exclusively via IDs signed by the server, so that no one can slip foreign instructions to the agent via a manipulated product description. Furthermore, the agent does not output raw HTML or Markdown but calls fixed, typed functions that render the interface in a controlled manner.
That is the difference between a chatbot that sounds plausible and an agent that a company can actually trust. The security measure isn’t found in an instruction that could be bypassed with the right trick. It’s embedded in the program code, which no prompt can alter.
Shopify and Shopware with Rapid Prototypes
Two of our technology partners have demonstrated in a very short time how a new standard can be used to quickly build an initial prototype, albeit starting from very different points.
On September 2, almost simultaneously with the Blueprint’s release, Shopify unveiled its own reference implementation. The shopping agent searches the catalog via the Universal Commerce Protocol, logs in via “Sign in with Shop,” and ultimately redirects the customer to the store’s regular checkout page. The agent does not process payment data at any point. The GitHub repository explicitly cites Anthropic’s Blueprint as its basis.
Shopware is applying the same principle in its own way: a public, UCP-native reference implementation that cryptographically signs all communication and allows merchants to specify which agents are permitted access. Here, too, the payment integration is, according to the company, not yet ready for production. So both partners are at the same stage: a prototype, not a finished product, but a clear signal of how seriously the topic is being taken in their respective platform roadmaps.
What this means for companies with extensive product catalogs
The Blueprint is not a product you can order tomorrow and put into operation the day after tomorrow. The demo environment for retailers runs on a fictional catalog of 87 products. The real effort begins afterward, when your own catalogs, pricing logic, and approval processes are integrated. For companies with large product ranges, many product variants, or products that require explanation, this is not a weekend project—regardless of the industry.
From our perspective, three points are more important than the question of whether you should already be launching a pilot:
First: Review your product data, not your agent strategy first. An agent can only suggest what its database allows. Incomplete attributes or free-text specifications hinder any standard, regardless of the provider.
Second: Define now what an agent is allowed to suggest and what a human must approve. The approval model from the blueprint becomes the standard. If you’ve already thought through where the line is drawn,you won’t waste time later on fundamental discussions.
Third: Start on the merchant side, not the customer side. An agent who checks inventory levels or provides a price proposal with a “before-and-after” comparison poses a significantly lower risk than one who speaks directly with your customers. This is precisely where the approval pattern can be practiced without any public risk.
Conclusion
Trust isn’t built by an AI sounding convincing. It’s built because it can’t make any decisions on its own when it comes to money or customer relationships. That’s exactly what Anthropic has codified with this blueprint, and with Shopify and Shopware, two of our technology partners are already demonstrating what the first prototypes look like.
The technology is therefore no longer an open-ended risk, but rather a matter of preparation. Those who get their product data and approval processes in order now will be ready when the first agent actually comes knocking on their door. Those who wait will have to address these issues later under time pressure.
*This blog post was created with AI assistance. The content has been editorially reviewed and revised.
About the author
Christian Schilling is the director of the business unit Complex Productsand leads the development of configuratorware, a configuration software program. He has nearly two decades of experience in digitalization, automation, and software development.
