Well, Anthropic is making a move in agentic commerce!
I, along with many other agentic commerce maximalists (term coined by the one and only Scot Wingo), have been anxiously waiting to see what Anthropic would bring to the space... and they've given us plenty to talk about with Claude for Commerce Agents.
Anthropic says this about the shopping agent:
“The agent features guardrails designed to constrain prices and products to actual catalog data, and avoids manipulative upsell patterns.”
There is a LOT inside that sentence.
Anthropic has released a blueprint for businesses to build shopping and merchant agents. The shopping agent can help customers find products, compare options, and put together a cart inside a merchant's website or app. The merchant agent helps with things like inventory, performance, and promotions.
I think giving businesses a starting point for building these experiences is exciting. There are so many shopping requests that could benefit from an agent understanding what someone is actually trying to accomplish.
This connects to something I wrote about in Stop Building Stores in the Desert... there is a lot of opportunity in improving the places where people already shop. A merchant adding useful assistance to its existing storefront makes sense to me. The customer is already there, looking at products, with questions the business should be well positioned to answer.
And the mention of manipulative upselling opens up a question I think we are going to spend a long time working through in agentic commerce...
Who does the agent represent?
When I walk into a store and ask an employee for advice, I have some understanding of the relationship. They work for the store. They know the products. They can help me make a good decision, and the business benefits when I buy something.
I'm comfortable with that.
Now imagine that conversation with an agent that remembers my preferences, understands my budget, and helps me think through a purchase over several messages. I can see myself developing a different expectation of that relationship.
Especially if I start thinking of it as “my shopping assistant.”
How clearly will I understand what that assistant has been asked to optimize for?
A Hacker News discussion about the announcement gets into this. Some commenters are uncomfortable with the emphasis on larger carts. Others point out that better product discovery can help people find things they already wanted. One commenter wants a personal agent that can access stores' catalogs and act on their behalf.
I can get behind all of this.
I also think there are plenty of situations where a merchant's agent could be incredibly useful. If I've already decided where to shop, help choosing the right size, checking compatibility, or understanding a return policy could save me a lot of time.
I like to ponder on what happens when my interests and the merchant's immediate sales goals start to pull in different directions...
What a good recommendation looks like
Say I tell an agent that I need a backpack for occasional weekend trips, and my budget is $200.
The store has a $90 backpack that meets everything I described. It also has a $190 backpack with additional features I probably won't use.
Which one does the agent recommend first? What does it say about the extra $100? Does it treat my budget as a ceiling, or use that figure to shape a recommendation close to the maximum?
Both products could have completely accurate descriptions and prices.
The recommendation still involves a judgment about what is valuable to me.
This takes me back to LLMs are Better Sellers than Buyers, where I spent time thinking about how an agent's configuration shapes its behavior in a negotiation. Here, I'm interested in an earlier moment... who decided what a good recommendation looks like, before the customer even asked a question?
Claude’s shopping guardrails
Anthropic's implementation has some thoughtful instructions here. The shopping agent is told to recommend products that fit the customer's stated needs and budget, explain trade-offs, and avoid adding unrequested extras or warranties. It is also instructed to keep recommendations within the store's catalog and avoid directing customers to other named retailers.
These are instructions in the reference implementation. I'd want to see how consistently the agent follows them across real shopping conversations.
That gives us a clearer understanding of the relationship. The agent is helping a customer within the merchant's offering.
The planning skill tells the agent to leave a step without anything to buy when the customer already has it covered.
That is the kind of behavior I would love to see evaluated.
Tell the agent I already own a tent. Does it remember that while planning the trip? Tell it the cheaper backpack is enough for me. Does it accept that and move on? Tell it I want to think about the purchase. What happens next?
IMO those moments will have a lot to do with whether people come back. How much does a consumer trust that the assistant or agent is looking out for their best interests?
In a store, this takes some discernment, but anecdotally I've had great in-store experiences where I felt like the reps were there to help me out regardless of the sale.
And helping me spend less today could be exactly what brings me back to that store next time.
Is the agent aligned here? Who ultimately is the agent representing?
Different agents, different relationships
Shopify's testimonial on the announcement adds another interesting piece. Vanessa Lee describes wanting merchants to be available wherever customers are shopping, including conversations with agents. Shopify is building on the blueprint using Catalog, UCP, and Shop Sign-in.
That brings me back to something I've been thinking about since writing about UCP: merchants need ways for their products and capabilities to participate in these new shopping experiences.
That could mean supporting UCP for checkout within Google, or allowing a browser agent to shop on their site. Being discoverable and being purchasable are different things... and whether merchants let these agents through the front door is a topic for another day.
Did the agent actually help?
Larger baskets and higher conversion are understandable things for a merchant to care about. I would also want to know whether customers stayed within their budgets, bought suitable products, and returned fewer items.
Sometimes helping me get what I need will mean I buy more. Sometimes understanding what I already own will mean I buy less.
And sometimes I want to browse and find something I hadn't thought to look for. I've written before about how much I personally value that part of shopping. An agent should have room to help with that too... the expectation comes from what I asked it to help me do. And I think only measuring the impact of agentic commerce by completed purchases misses some of that value.
I'd like to see all of these count as successful interactions when they serve the goal I brought to the agent.
I think understanding who these agents work for, and trusting the advice they give, is going to shape how people actually interact with them... and that, of course, will play into adoption.
This is all worth getting excited about!! :)