uygnoey Commerce agent blueprint
Anthropic publishes a blueprint for building commerce agents at github.com/anthropics/commerce-agents, with reference implementations for a shopping agent and a merchant agent across retail, travel, telecom and ticketing.
Reported results from retailers running shopping agents on Claude: carts up to 35% larger, and shoppers 60% more likely to complete a purchase.
Instructions
1. Pick which agent you are building first
The blueprint ships two, and they have different owners inside most companies.
- Shopping agent — consumer-facing. Search, comparison, multi-item assembly, preferences, cart building, and customer service questions.
- Merchant agent — business-facing. Sales performance, inventory, pricing and promotions, campaign drafting.
Full capability lists are in references/blueprint-contents.md.
2. Map the integration points against systems you already have
The shopping agent embeds in an application or website and needs integration points for:
- catalog search
- cart management
- checkout
- customer preferences
- order history
Scope the project by which of these already exist. The agent calls them; it does not replace them.
3. Keep the guardrails that ship with it
Two guardrails come with the reference shopping agent and are worth treating as non-negotiable:
- Prices and products are constrained to catalog data. The agent does not state a price or an attribute it did not retrieve.
- No manipulative upsell patterns.
And one for the merchant side:
- Agents suggest changes; humans approve before deployment.
4. Choose where it runs
The same deployment can run on the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. See references/deployment-options.md.
5. Fork, then read the engineering guide
- Fork github.com/anthropics/commerce-agents.
- Read the engineering deep-dive at claude.com/blog/the-anatomy-of-effective-commerce-agents for the architecture, latency, caching, memory, safety and eval practices behind the blueprint.
- See vertical demos and request a working session at claude.com/solutions/commerce.
Teams report the setup is fast: engineers at Wix had working commerce agents within fifteen minutes; at Fetch, both agents ran locally in under an hour with live conversations working on the first attempt.
Examples
Scoping a retail deployment
A retailer has catalog search and an order system, but no stored customer preferences. The shopping agent still works — search, comparison, cart building and order tracking are all available. Personalization is the gap, and it is a separate piece of work (see the memory guidance in the engineering deep-dive), not a blocker for a first deployment.
Deciding which agent goes first
A marketplace wants both. The merchant agent has a smaller blast radius — it suggests and a human approves, so a bad suggestion costs a review rather than a customer — which makes it the safer first deployment while the shopping agent's guardrails and evals are built out.
A guardrail question
"Can the agent offer a discount to close a hesitant shopper?"
Not by inventing one. Prices are constrained to catalog data, and manufactured urgency is explicitly out of scope. A promotion the merchant has actually created, retrieved from the catalog, is fine.
Source
- https://claude.com/blog/claude-for-commerce-agents (published 2026-09-02)