Claude Commerce Agents: Ship AI Shopping in Days (Open Source)

Anthropic open-sourced two production-grade AI agents on September 2 that developers can clone, configure, and run as a live demo in under fifteen minutes. The anthropics/commerce-agents repo — Apache 2.0, no vendor lock-in — ships a shopping agent for customer-facing storefronts and a merchant agent for back-office operations, with reference implementations across retail, travel, telecom, and entertainment. Shopify and Priceline are already running production builds on Claude. What remains to be seen is whether this blueprint solves real developer problems or is a well-packaged way to accelerate Claude API spend.

What Is Actually in the Repo

Two agents, seven pip packages, four industry verticals. The shopping agent handles customer-facing interactions: catalog searches, multi-item assembly, cart management, order inquiries, and personalization across five modular skills. Importantly, it does not complete purchases — it hands off to the retailer’s existing checkout. The merchant agent handles back-office operations: sales analytics, inventory alerts, listing updates, pricing recommendations, and campaign drafting. Every merchant-agent write is staged for human approval before execution.

Getting started takes three commands:

git clone https://github.com/anthropics/commerce-agents.git && cd commerce-agents
pip install -r requirements.txt && cp .env.example .env
python scripts/run_demo.py retail

Add your ANTHROPIC_API_KEY to .env and you get a live storefront API on port 8000 and a React frontend on port 3000. Wix engineers reportedly had a working agent taking prompts in fifteen minutes. Zomato and Fetch both got it running locally in under an hour. Those numbers may be cherry-picked for the press release, but the quick start is genuinely minimal. Three runtime modes are available: the Messages API reference implementation, the Claude Agent SDK for managed loops, and Claude Managed Agents (currently in beta). Deployment targets include the Claude API directly, Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI.

The Architecture Decision Worth Stealing

The most valuable part of this blueprint is not the commerce code — it is the architecture reasoning behind it. Anthropic explicitly tested three approaches before settling on their design:

• One giant system prompt for everything
• Specialized subagents with handoffs between them
• A single agent equipped with modular skills

The single-agent-with-skills approach won across enterprise deployments on quality, cost, and latency. The reasoning is sound: a commerce conversation is a tightly-coupled session that spans multiple intents across many turns. Subagent handoffs are state-lossy — each one can cost several times the tokens and add seconds of latency. Skills load into the agent that already holds the full conversation history, delivering the same modularity with none of the handoff overhead.

This is not just a commerce insight. Any agent task where context needs to survive across intent shifts benefits from this pattern. Subagents remain the right call when a task genuinely splits into clearly separated work with different tool needs — a codebase refactor using parallel backend, frontend, test, and review agents, for instance. But if your “multi-agent” design is really just one conversation being shuttled between specialists, you are paying a latency and quality tax for no gain.

Safety Lives in the Harness, Not the Prompt

Anthropic’s safety architecture addresses the most obvious production failure modes. The key principle: safety enforcement belongs in the harness code, not in prompt instructions the model could ignore or misinterpret under adversarial input.

Concrete mechanisms in the blueprint include ID validation (only server-issued IDs accepted — hallucinated or user-pasted IDs are refused before hitting the backend), staging before execution (no tool call moves money directly; everything is staged with server-generated IDs), and protected content handling (fees, disclosures, and regulated copy come from approved strings the model cannot paraphrase). Transaction limits are enforced on resulting state, not the request, preventing parallel tool calls from combining to exceed caps. All untrusted input — product listings, reviews, seller messages — runs through a sanitizer that strips prompt injection attempts before the model ever sees them.

HackerNews developers flagged real concerns about agentic checkout. Prompt injection in commerce contexts is a live threat. The blueprint’s harness-first approach is the right response, but it requires developers to actually implement the safety patterns rather than treating the demo as production-ready.

The Numbers and the Caveat

Anthropic claims cart sizes grew 30 to 35 percent and customers were 60 percent more likely to complete a purchase for a partner running Claude-based commerce agents. These are compelling numbers — and they come from Anthropic’s own reporting on a single unnamed partner. They are not an independent benchmark.

Context matters here. OpenAI retired its in-chat Instant Checkout in March 2026 after Walmart found conversion inside the chatbot ran roughly three times below their own site. The graveyard of agentic checkout features is not empty. Anthropic’s open-source approach and deliberate decision not to complete purchases are meaningful differentiators, but the performance claims deserve scrutiny before anyone rewrites their checkout funnel around them. The Visa and Mastercard partnership adds credibility to the enterprise story, but enterprise validation and developer-facing ROI are different things.

What Developers Should Do Now

Two backend interfaces need implementing to connect the blueprint to real systems:

StorefrontBackend: search(), get_product(), get_cart(), add_to_cart(), checkout(), get_orders(), get_policies()

MerchantBackend: get_analytics(), get_catalog(), update_listing(), get_inventory(), manage_inventory(), manage_pricing(), manage_campaigns()

The Claude Code plugin scaffolds the integration:

claude plugin install commerce-builder@claude-commerce-agents
/scaffold-commerce-agent a shopping assistant for our store

One important caveat before you fork: Anthropic stated explicitly that this is a reference implementation, not a product. They will not maintain it or accept outside contributions. Treat it as a starting point your team will own, not a dependency you can update from upstream.

The blueprint compresses months of agent architecture trial-and-error into a repository. The underlying technical decisions — skills over subagents, safety in the harness, prompt caching structured for 90 to 99 percent hit rates — are defensible and transferable well beyond commerce. Start with the retail vertical, implement the two backend interfaces, and run the smoke tests before pointing it at production traffic.