What Anthropic actually released
In September 2026, Anthropic published Claude for commerce, an open blueprint for building commerce agents. It is not a hosted ecommerce platform, marketplace or checkout product. Anthropic describes it as reference code and patterns that a team can fork, connect to its own systems and maintain.
The blueprint includes working patterns for two broad agent types: a shopping agent for customers and a merchant agent for the people operating the store. Anthropic provides implementations across the Messages API, Claude Agent SDK and Claude Managed Agents, plus a Claude Code plugin to scaffold the design against a business's own catalog and services.
Shopping agent vs merchant agent
Customer-facing shopping agent
A shopping agent sits on the buyer side of the journey. It can translate conversational intent into catalog actions: finding products, comparing options, building a list or cart, answering approved policy questions and handing the shopper into checkout.
The strongest use cases are where the shopper's need is difficult to express through filters alone. “Build me a home-office setup under $800 for a small apartment” contains budget, space and compatibility constraints that an agent can reason across—provided every recommendation is grounded in real catalog data.
Internal merchant agent
A merchant agent serves store operators. It can answer questions about sales or stock, surface exceptions, draft promotions, prepare product-content changes and stage operational actions. The key word is stage: a sensible production design separates analysis and drafting from consequential writes.
These two agents should not share the same permissions. A shopper should not receive internal margin data, and a merchant assistant should not have unrestricted authority to change prices, issue refunds or publish catalog edits without policy controls.
How the architecture works
Anthropic’s engineering guidance uses a relatively simple pattern: one model in an agent loop, with skills for specialised behaviour and tools that call the systems the business already runs. The model decides which approved tool to use; the tool performs the deterministic action.
A production architecture commonly includes:
- Identity and session context so the agent knows which shopper or merchant context is permitted.
- Catalog tools for search, product details, variants, pricing and availability.
- Cart and checkout tools that rely on the ecommerce platform rather than asking the model to simulate a transaction.
- Policy and support tools grounded in current returns, shipping and customer-service information.
- Memory only where the business has a clear privacy and consent model.
- Evals and logs so the team can reproduce failures and test changes before release.
For BAGAI, the important design principle is that the agent should reason over trusted tools instead of receiving broad database access. That keeps the model replaceable and makes permissions easier to audit.
Where ecommerce businesses can use Claude Commerce patterns
1. Guided product discovery
Translate a broad customer need into structured criteria, query the catalog, compare options and explain trade-offs. This is particularly useful for large catalogs, complex products, bundles or purchases where compatibility matters.
2. Shopping lists and bundles
Create multi-item plans while checking budget and product constraints. The agent can suggest a basket, but availability, price and variant validity should come from live tools.
3. Customer care inside the buying journey
Answer shipping, returns, sizing or product questions without forcing the shopper into a separate support surface. High-risk account or refund actions should still use authenticated workflows and authority limits.
4. Merchant analytics
Let operators ask questions such as which products are slowing, where inventory is tight or which campaign changed demand. The agent can combine structured commerce data with plain-language explanation, then link back to the underlying numbers.
5. Catalog and promotion assistance
Draft product copy, promotion ideas or merchandising changes. For production use, BAGAI would normally stage those changes for review rather than letting generated content publish immediately.
Safety and control are part of the product design
Commerce agents can affect money, inventory, customer trust and legal obligations. Prompt instructions alone are not enough. The surrounding harness should enforce what the agent may read and write, which prices are valid, how many items may be changed, when a refund or discount requires approval and when to escalate to a person.
Anthropic’s reference material emphasises grounded pricing and availability, hard limits on cart/refund authority and human escalation. WooCommerce’s experimental adaptation similarly stages merchant-side changes behind approval. These are useful patterns because the agent can remain helpful without becoming the final authority for every action.
How to measure whether an AI commerce agent is working
Do not measure success by conversation volume. Connect agent events to the ecommerce funnel.
| Stage | Useful measures |
|---|---|
| Discovery | Agent opens, query completion, product-result engagement |
| Consideration | Product comparisons, detail views, assisted add-to-cart rate |
| Checkout | Cart handoff, checkout start, purchase conversion |
| Commercial | Average order value, revenue per assisted session, repeat purchase |
| Operations | Merchant time saved, exception resolution, approval rate, error/rework |
For a BAGAI implementation, GA4 can capture the behavioural layer while the ecommerce platform remains the source of truth for orders and revenue. A unified dashboard can then compare agent-assisted and non-agent journeys without pretending the agent caused every sale it touched.
A sensible pilot plan
- Choose one high-friction customer or merchant workflow.
- Define the trusted systems and tools required to complete it.
- Start read-only where possible, then add narrowly scoped actions.
- Write eval cases from real product, policy and edge-case scenarios.
- Instrument the flow before launch so conversion and failures are measurable.
- Run in staging, review logs and add explicit escalation conditions.
- Expand only after the bounded flow proves useful and dependable.
This avoids the common mistake of starting with “build an autonomous ecommerce agent” instead of starting with a measurable job that an agent can improve.
Primary sources
This guide is based on current first-party documentation and announcements:
- Anthropic: Claude for commerce
- Anthropic: Building commerce agents with Claude
- Claude Platform Docs: Commerce agent guide
- Anthropic: Anatomy of effective commerce agents
BAGAI is independent and is not affiliated with Anthropic.