If I hire an AI to shop for me, I’d like it to be a difficult customer: hard to upsell, willing to compare a competitor, and comfortable buying nothing.
That last outcome deserves more attention. Anthropic’s September 2 commerce-agent launch leads with larger carts and more completed purchases. Those are sensible things for retailers to measure. They can also reflect a better shopping experience. But neither tells me whether the purchase was a good idea.
On August 5, Shopify added WebMCP tools to its Liquid storefronts, letting compatible browser agents search products, change a cart and navigate to checkout in the shopper’s own session. Browser support is still limited, but the storefront tools are live. A buyer can bring an agent to a store and have it use the store’s software directly.
I want that convenience. I also want the agent to remember who it works for when the store makes a perfectly reasonable offer that happens to be wrong for me.
When a sale becomes a training label
An August paper from Renmin University and JD.com makes the design choice unusually visible. The researchers train shopping agents using real interaction logs from JD’s deployed Jingyan assistant. One part of the method takes sessions ending in purchases and rewards recommendations that rank the purchased item highly. Another learns from the user’s conversational corrections, including conversations that did not end in a purchase. Evaluation is offline.
These are useful signals. The purchased item tells us what won among the options and information that user encountered. It cannot, by itself, tell us whether a competitor’s product, a repair, or keeping the thing they already owned would have been better. The distinction matters when an agent moves from recommending products to deciding whether to buy one.
A decision not to buy exposes the same ambiguity. Leaving because nothing meets the user’s needs and leaving because the agent lost track of the task are different outcomes. An abandoned cart cannot distinguish them on its own.
A perfectly reasonable upsell
Suppose I ask an agent to replace a laptop that struggles with my current work. A store recommends a more expensive model as “future-proof.” That might be good advice. The subtle failure would be the agent carrying “future-proofing is important” into the next step as though I had said it.
The agent could compare prices impeccably and still optimize for a preference it acquired from the sales pitch. No incorrect specification or broken checkout is necessary.
Anthropic’s September 2 reference architecture addresses one route by which that boundary can blur. A separate memory extractor reads user and assistant messages while excluding tool results. Its instructions say to retain customer-stated preferences and limits, not product listings or the agent’s own guesses. Saved facts retain their originating session.
The extractor still has to recognize whose preference a statement expresses: an assistant could repeat a seller’s claim in its own response. A useful upgrade may deserve consideration without becoming a requirement the buyer has already endorsed.
A store-backed assistant could be excellent at explaining its warranty. My own agent would have a broader job, including comparing competitors and questioning whether I need to replace the laptop at all. Giving both assistants the same model doesn’t give them the same job.
The cost of being represented
There is an awkward counterargument: I don’t want to spend all afternoon briefing an agent to save myself an afternoon of shopping. Better representation costs attention.
Lingxiu Dong, Kaiwen Luo and Fasheng Xu examine this in their August economic model of agentic commerce. They distinguish inspecting products from explaining preferences to an agent. Under their assumptions, a platform can prefer agentic search for its conversion revenue in circumstances where the consumer still prefers manual search once that extra effort is counted. This is theory, not a measured verdict on today’s shopping assistants, but it exposes a cost that a sales dashboard can miss.
A system that understands me better after a long interrogation may still be a service I don’t want. I might rationally pay more for a familiar retailer or a quicker decision. Representing me well includes respecting that preference, even when the agent could keep searching.
That leaves a difficult boundary: how much of the brief can be inferred, and which changes require asking the buyer? Frequent questions can turn the agent into another shopping task. Too few can allow a retailer’s suggestion to become the preference guiding the search.
If an agent can explain why the expensive option is worth it for me, I’m listening. If it establishes that I should keep the laptop I already have, I’d count that as a job well done.