Briefing · Commerce agents
AI shopping assistants are getting easier to build. Connecting one to your business is still the work.
On 2 September 2026 Anthropic released a blueprint for commerce agents: a shopper assistant, a merchant assistant, examples across four sectors and a Claude Code plugin (an add-on for Anthropic’s coding tool). It makes the conversation easier to build. Your prices, availability, terms and approvals still have to be connected, and the promise at the end is yours to keep.
Published
Anthropic, 2 September 2026
Two assistants for two different people
Anthropic’s blueprint gives engineering teams a starting point they can adapt to their own systems. It comes with two reference assistants.
For the shopper
- Asks for suitable products
- Compares the options
- Puts together a cart
For the store operator
- Looks into how the business is doing
- Reviews stock
- Prepares a promotion
Anthropic also names companies already building commerce experiences with Claude. Those wider deployments aren’t the examples in this repository.
Reading the examples
Useful reference, not a receipt
The repository uses fictional businesses, products and people. It doesn’t place real orders, charge cards or change live listings.
What it demonstrates
- A shopper agent that compares options and builds a cart
- A merchant agent that reviews stock and prepares a promotion
- A checkout tool that hands the cart to the business’s own app to complete
- Worked examples across four sectors, with fictional companies
What it doesn’t establish
- An order placed at a real merchant
- A card charged
- A live listing changed
- That a given deployment enforces approval in the app
It works the other way too. Anthropic says its work draws on agents running in production, so the fictional examples don’t mean every Claude commerce deployment is a demo. Evidence about a real merchant has to come from that deployment.
Connecting it to a business
Five steps that have to hold for a real business
The repository’s integration guide asks whoever deploys it to connect the agent to the business systems behind it.
Know who’s asking
The application identifies the caller, including a guest, and limits what that session can see and do.
Read your real records
Your systems supply the catalogue, orders and policies for that person. Missing figures stay missing rather than turning into a convenient zero.
Propose the action
A cart, or a change staged for approval. Nothing is final yet.
Approve it in the app
The application’s own check, not the model’s. Limits are checked again when the change is applied.
Complete it once
Checkout has its own completion path. Multi-step actions need the right order and protection against recording the same order twice.
The reference shows how these steps can work with fictional data. Making each one hold for a real business, approval included, is the deployer’s work.
Why approval has to belong to the application
In Anthropic’s engineering guide, a proposed merchant change gets an identifier and reaches an approval screen, and limits are checked again when it is applied. That matters when stock, permissions or policy can change between a suggestion and its execution.
The reference safety documentation sets host approval as the default, but a developer can change the policy. A convincing confirmation in the conversation is therefore not enough evidence that a particular installation enforces the right approval. Test it in the app people will actually use.
What it means at your door
Which location, which table, which time, which terms?
Illustrative example. It is not a claim about Anthropic’s restaurant integrations.
For a local business, a useful answer has to survive that detail. Take a restaurant handling a group booking.
A proposed time isn’t a held table
Until your booking system holds it, someone else can take it.
A deposit shown isn’t a deposit paid
The chat can display an amount; only your payment record shows it arrived.
A payment doesn’t set the terms
A payment confirmation doesn’t establish the cancellation terms on its own. You still have to state them.
Commerce platforms
Three platforms, three different stages
Anthropic’s announcement includes statements from three platforms. They describe different work at different stages, not one integration every merchant can switch on.
Shopify
- Describes a reference storefront connected to a merchant’s store.
Wix
- Describes a pilot.
Square
- Describes agentic tools it is building for sellers.
Our read: platforms can make their existing merchant relationships more useful through agents. The test is whether a merchant can connect the right records, grant limited authority and see afterwards what happened. A quick setup demo answers only part of that.
What you can prepare now
Start with one job you can stand behind
Pick one small, well-defined job
A group booking or a promotion, say, before any grand promise of automation.
List the records it needs
For a booking, availability and terms; for a promotion, eligible stock and price limits.
Name who approves
Write down the person allowed to approve each change, and test it in the app.
Define what counts as done
A held table, a paid deposit, a fulfilled order: decide which result completes the job.
Plan for things going wrong
Say what happens when an item is unavailable, a price changes, permission is missing or the same request arrives twice.
Keep proposed, approved and completed apart
Show them as three different states, both on screen and in the record of what happened.
Keep reading
Related Briefings and next steps
Three related Briefings, and where to start with your own business.
- For eligible US restaurants on Square, a question in Maps can now end as an order in the kitchenThe restaurant still owns the records and the promises behind the order.Read the Briefing
- AI assistants are moving from recommending to bookingThe wider shift: AI surfaces are starting to act for a customer, not just answer.Read the Briefing
- The committability ladderWhere each step from finding a business to paying it sits, and which step only the business can take.Read the Briefing
- Merchant InfrastructureThe business facts and commitments AI agents depend on, kept accurate at the source.See how it works
- Check your AI PresenceSee what AI answers say about your business today.See the check
Sources
Where this comes from
Anthropic’s announcement, its public pages and its reference repository at a pinned commit. This Briefing does not establish an Obenan integration with Anthropic’s commerce blueprint, a payment service, or a completed merchant transaction. Obenan has no partnership, endorsement, integration, pilot or special-access relationship with Anthropic.
- Anthropic’s commerce agents announcement
Anthropic Published Checked
The release, the companies Anthropic names and the platform statements.
- Reference repository README, pinned commit
Anthropic on GitHub Checked
What the examples contain and what they don’t do.
- The anatomy of effective commerce agents
Anthropic Checked
Production experience and approval outside the model.
- Reference safety model, pinned commit
Anthropic on GitHub Checked
Default host approval, and that a developer can change it.