Anthropic's Commerce Agent Blueprint: Two Reference Agents You Can Run in an Afternoon

On September 2, 2026, Anthropic published a commerce agent blueprint that changes how teams should think about building AI shopping experiences. The release is at github.com/anthropics/commerce-agents, Apache-2.0 licensed, and it contains two working reference implementations: a customer-facing shopping agent and a staff-facing merchant agent.

The blueprint is not a model release or a platform feature. It is a product-development framework — the harnesses, patterns, and guardrails an engineering team needs to get a commerce agent running in days instead of months.

What Shipped

Two reference agents form the core:

Shopping Agent (customer-facing): Handles product discovery, comparison, cart management, and checkout flows. It runs in the storefront, talks to your catalog API, and enforces guardrails like price limits, inventory checks, and policy compliance.

Merchant Agent (staff-facing): Runs in the back office. It handles inventory updates, order reconciliation, refund approvals, and catalog maintenance. It talks to your ERP or admin APIs and enforces the same policy layer.

Both agents share a common architecture: an agent loop, a tool layer over your catalog/admin APIs, an approval gate for sensitive actions, and an eval suite. Anthropic built this scaffolding so teams don't have to rebuild it.

Four Vertical Implementations

The repo includes complete example implementations for:

• Retail: Product search, variant selection, size/color logic, bundle recommendations
• Travel: Flight/hotel search, date flexibility, loyalty tier logic, cancellation policies
• Telecom: Plan comparison, device upgrade eligibility, port-in flows, promo stacking rules
• Ticketing: Seat maps, dynamic pricing tiers, transfer/refund rules, waitlist management

Each vertical demonstrates how the same agent skeleton adapts to different catalog shapes and business rules.

The Guardrails Layer

The most valuable part of the blueprint is the guardrails system. It sits between the agent and your APIs and enforces:

• Price floors/ceilings: The agent cannot quote below cost or above MAP
• Inventory atomicity: Cart holds are reserved, not just checked
• Policy compliance: Refund windows, transfer rules, promo exclusions
• Audit logging: Every agent action is traceable to a session and a decision

This is what separates a demo from a production agent. Most teams build the agent loop first and bolt on guardrails later. The blueprint inverts that order.

Claude Code Plugin

The release includes a Claude Code plugin that lets you develop and test the agents locally. You can spin up the shopping agent against a mock catalog, run the eval suite, and iterate on prompts without deploying. The plugin handles the MCP connection to your local catalog API and the approval gate simulation.

This means you can validate the agent behavior against your actual catalog data and business rules before any production deployment. The eval suite includes regression tests for the guardrails, so a change to a price rule or inventory policy immediately shows whether it broke anything.

What This Means for Builders

If you are building an AI shopping assistant or a merchant tool today, you were going to build this scaffolding yourself. The agent loop, the tool layer, the approval gate, the eval suite — every team rebuilds them. Anthropic has now released that scaffolding as code.

The strategic signal: Anthropic is moving from "here is a model" to "here is how you ship a product on the model." The blueprint is a reference implementation of the patterns they see working in production with partners like Shopify and Priceline (both run live agents on this technology).

The code is Apache-2.0. You can fork it, strip the verticals you don't need, wire your own catalog and ERP, and have a production-grade agent skeleton in an afternoon. The eval suite tells you when it's ready to ship.

Practical Takeaway

1. Clone the repo: git clone https://github.com/anthropics/commerce-agents
2. Run the retail demo against the mock catalog
3. Read the guardrails module — that is the pattern worth adopting
4. Wire your catalog API, configure your price and inventory rules
5. Run the eval suite. When it passes, you have a shippable agent

If your team is evaluating whether to build or buy an AI commerce layer, this blueprint is the benchmark. It shows what production-grade looks like when the model provider builds the reference implementation.