Anthropic released AI sales agent code

Anthropic revolutionizes commerce with open-source AI agents

On September 2, 2026, Anthropic took a bold step by publishing a comprehensive blueprint for commerce agents, released under the Apache 2.0 license. This initiative includes not only two functional AI agents but also examples for four key verticals: retail, travel, telecom, and ticketing. Additionally, a Claude Code plugin is provided, designed to facilitate adaptation to any product catalog.

Two AI agents to transform your business

Anthropic has developed two main types of agents, each with specific functions to optimize commercial operations:

1. Buyer Agent: This agent is designed to interact directly with customers within your app or website. Its capabilities are extensive: it searches for products in the catalog, builds multi-item orders, remembers user preferences, displays product comparisons within the conversation, constructs the shopping cart, and prepares it for checkout. The implementation of the payment process is left to the company, ensuring flexibility and control.

2. Store Agent: Geared towards company staff, this agent is an invaluable support tool. It can explain sales performance, monitor inventory, recommend prices and promotions, and draft marketing campaigns. It's important to note that, to ensure security and human control, no changes suggested by this agent are applied without a person's final approval. This highlights an approach to intelligent assistance without replacing critical oversight.

Instruction Injection Defense: A Crucial Detail

A little-publicized but fundamental aspect lies in these agents' ability to ignore malicious or unintended instructions. The official screenshot of the store panel, published by Anthropic, reveals a scenario where the AI assistant identified and treated as message content (not as executable commands) two buyer messages that included directives to refund and change prices. This ability to differentiate content from direct commands is a robust defense against instruction injection, highlighting the maturity and security of Anthropic's design.

Understanding the “fine print” of open source

It is crucial to understand the scope of this code release. While it is a valuable initiative, it is a reference implementation. This means it does not receive continuous maintenance or contributions from Anthropic, does not execute real orders or manage payments, does not modify live listings, and the examples do not include authentication. Business rules and authorization are the sole responsibility of whoever deploys the solution. The figures mentioned, such as “up to 35%” larger carts and a “60%” higher probability of completing the purchase, come from retailers already using agents with Claude, not from an independent third party. This underscores the need for internal evaluation to measure the actual impact.

How to apply it in your business?

Anthropic's blueprint release represents a strategic opportunity for companies looking to innovate in commerce automation. Consider:

• Customization: Adapt the Claude Code plugin to your specific catalog and business workflows. The flexibility of open source allows for deep integration with your existing systems.
• Operational Efficiency: Use the store agent to assist your staff, optimizing inventory management, pricing, and campaigns, always with human supervision.
• Improved Customer Experience: Implement the buyer agent on your website or app to offer a personalized and fluid shopping experience, from search to cart preparation.
• Internal Development: Leverage the Apache 2.0 license to build upon this foundation, developing customized solutions that fit your unique market needs, maintaining control over system evolution and data security. This foundation can be a starting point for your development team, avoiding starting from scratch in creating conversational agents for commerce.