Anthropic Releases Reference Implementation for Commerce Agents

Anthropic has released 'Claude Commerce Agents', a reference implementation that enforces permissions (such as cart, inventory, and pricing) and approval boundaries via code when operating conversational agents. Unlike previous examples that relied on prompt instructions and failed to resolve permission issues, this repository clearly separates rules that are enforced by code from those left to model judgment.

Core Architecture and Execution Model

The repository includes two roles: a customer-facing shopping agent and an employee-facing seller agent. By defining prompts, skills, tool contracts, and gates only once, they are executed identically across three execution paths (Messages API, Claude Agent SDK, and Managed Agents). While the scope of rule application varies by path, rules enforced within tool calls are applied consistently regardless of the path.

Safeguards and Permission Management

For stability, rules are separated into three layers.

• Enforced by code: External text isolation, provenance verification (allowing only tool return identifiers within the session), absence of payment code, mandatory human approval (staged change), and memory write restrictions, among others
• Requested via prompt: Treating isolated text as a reporting subject rather than an instruction, limiting numerical citations, among others
• Deployer responsibility: Authentication, authorization, request rate limiting, fraud detection, among others

In particular, all write operations by the seller agent are designed to accumulate as staged changes requiring human approval rather than being applied immediately, thereby reducing the risk of malfunction.

Usage and Limitations

Demos can be run based on virtual data from four industries: retail, travel, telecommunications, and entertainment. It supports reviewing existing agents or generating new agents via the Claude Code plugin. However, as this is a reference implementation rather than a finished product, authentication and payment logic must be implemented directly, and the project does not accept long-term maintenance or contributions.