Deep Dive — Walmart Agent Gateway: AI agents shop, you set the limit

The question people will actually type is whether an AI agent can shop at Walmart for them. From next month the answer is yes: Walmart will let you connect an agent to Walmart.com and Sam's Club, give that agent its own profile inside your account, and cap what it can spend. The retailer announced the system on October 21, in a post on its corporate newsroom, and it is the first time a major American retailer has published rules for agents rather than a wall against them.

The wall is the comparison that matters. Amazon has spent a year arguing that a site gets to decide which software walks in, and in September it cut off Meta's Muse shopping agent mid-crawl — a fight we went through in Amazon bars Meta's Muse agent from shopping its site. Walmart is taking the opposite bet: come in, but through my door, on my terms, with a credential I issued.

What Walmart actually shipped

Three pieces, and only one of them is a product you can use today.

The Agent Gateway is a server that speaks the Model Context Protocol, the open standard Anthropic introduced for connecting models to outside tools. Instead of an agent loading Walmart.com in a browser and clicking through a checkout designed for humans, the gateway exposes catalog search and purchasing as structured calls. Walmart says the browser path "can be slow and error-prone, and agents may get stuck along the way," which is an accurate description of what agentic shopping looks like in practice. The gateway starts as a beta for companies and developers, per TechCrunch, at a signup page on walmart.com — no consumer date attached.

There is a small tell in the announcement: Walmart's own copy calls MCP the "model text protocol." It is the Model Context Protocol. When the retailer writing the terms cannot name the protocol correctly, you are reading a document written in the first weeks of a transition, not at the end of one.

Agent Profiles arrive "in the next month," which puts them in November. The first time you use an agent through the gateway, Walmart creates a profile for that specific agent inside your account. You link it by signing into Walmart, choose how purchases get approved, and can delete the profile at any time. Purchases made by the agent show up separately from yours. Walmart's analogy is a second cardholder on a credit card: same account, its own identity, a limit you set.

Terms for AI agents are the part nobody will read and the part that decides how this goes. Walmart lists six conditions. Customers approve agent purchases by default. Agents must not act beyond what the shopper intends. Agents never receive card numbers and cannot add or change a payment method. They can reach only the information the task needs. The agent company's privacy policy governs what the agent sees. And agents that break the terms get blocked from the site.

How it works, in practice

You start inside the agent's app, not Walmart's. The agent calls the gateway, Walmart spins up a profile, and you sign in once to authorize it. Then you pick controls: approve every purchase, or hand the agent a spending limit — Walmart's examples are fifty, one hundred and two hundred dollars — and watch its orders in their own tab. A Walmart+ membership is not required.

The detail that will decide whether people trust this is the returns line. Walmart says its returns process applies, "including if an agent doesn't act as intended." That is the most consequential sentence in the announcement, because it moves the cost of a bad agent decision off the customer and onto the merchant. A shopper trying an agent for the first time is not risking forty dollars of wrong dog food; they are risking a trip to the returns desk. Walmart is underwriting the experiment so the customer will run it.

Then there is the fine print inside the fine print: Walmart "may require additional approvals even for purchases within an agent's spending limits." A cap is not a guarantee. The merchant keeps a switch that can stop any transaction for any reason, which is the same veto Amazon exercises at the front door, moved to the moment before payment.

Why a gateway beats a browser

Walmart's stated reason for the gateway is reliability, and that is real. Browser agents break on cookie banners, sign-in walls, address forms and anything that assumes a human finger. A protocol call either returns a product or an error, and the error is legible, which is what makes retries cheap.

The unstated reason is data. On the open web, an agent that logs into your account can see your order history, your saved payment methods and your addresses. Through the gateway, Walmart can answer "find this customer's usuals" without handing over the history that answer is derived from. Walmart's example is sourdough bread and a particular brand of milk: the agent asks for the usuals, Walmart returns items, and the agent never learns why.

That is a scoping decision, and it is the right one. It is also a moat. The customer's shopping memory stays with Walmart, which means the agent's personalization is only as good as the merchant chooses to make it. Flip it around and you can see what Walmart is protecting against: an assistant that learns your grocery list, your brand preferences and your replenishment cycle across every retailer is an assistant that can route your next order to whoever pays it. Walmart is willing to be a supplier to that assistant. It is not willing to hand over the customer file that makes the assistant valuable.

The price line is aimed at someone

"We price products, not people," Walmart writes, "so the prices that an AI agent sees are the same prices that a human would see on Walmart.com, with applicable offers and fees shown before purchase."

Read that as a commitment, because it is one, and then read it as a warning. The obvious business model for agentic commerce is differential treatment: a fee for the agent channel, a discount for customers who let an assistant negotiate, a sponsored slot for the product the agent surfaces first. Walmart has ruled out the price half of that in public. It has said nothing about the ranking half. Walmart runs a large advertising business, and the announcement does not describe how products are ordered inside gateway results or whether promoted items can appear there. Price parity without ranking transparency is half a promise.

This is where Walmart's model diverges from the two alternatives on the table. ChatGPT's commerce push is being built on an ad business that reached 100,000 advertisers and a billion-dollar run rate inside 200 days, as we covered in ChatGPT ads: 100,000 advertisers and a $1B run rate in 200 days — the assistant gets paid to influence the choice. Meta's Muse has been pitched as taking a small fee on transactions instead of selling attention. Walmart's version keeps both the transaction and the relationship, and pays the agent nothing it has disclosed.

Walmart, Meta and Stripe already published a shared protocol for exactly this problem — the plumbing we went through in Meta, Walmart and Stripe publish the Personal Agent Protocol. The gateway is that idea with a retailer's name on the contract. It is less a door thrown open than a customs post: the agent may enter, it may carry only declared goods, and the officer at the desk decides.

The 57 percent problem

Here is the part the announcement cannot solve with policy. Shopping agents are not yet good at the job Walmart is inviting them to do.

EComAgentBench, a benchmark built on 662 tasks drawn from real Amazon products and reviews, splits each shopper's requirements across three places: what they said, what sits in a tool-gated profile, and what only comes out if the agent asks the right question. Agents get 100 tool calls to converge on one product. Of the seven models tested, the strongest reached 57.1 percent overall accuracy, and scores fell as requirements moved from the visible query into the hidden profile, according to the benchmark's authors. A separate September benchmark, RealWorldShop, built on 3.28 million products, found systems producing locally plausible answers while failing at state tracking and at updating constraints mid-conversation.

Walmart's "usuals" feature is, mechanically, a tool-gated profile. That is precisely the condition under which measured agent accuracy drops. So the design is correct and the capability is unproven at the same time, which is a strange place to launch a checkout.

The returns policy is what makes that acceptable, and it is also the bill. If agents misbuy at the rates those benchmarks imply, Walmart absorbs the reverse logistics: pickups, restocking, shrink on perishables that cannot be resold. Retailers already eat return fraud from humans, and the industry has spent a decade tightening return policies because of it. Walmart has just written a policy that says, in effect, agent errors are our problem. That is a bold promise to make before you know the error rate.

What Walmart did not say

No volume. No count of approved agents. No transaction figures. No date for general availability beyond a developer beta and "next month" for profiles. No statement on whether the gateway charges agents, merchants or nobody. No published list of which assistants will be allowed in, which matters because "approved agents" is the whole ballgame — an agent that is not approved does not exist on Walmart.com, whatever the customer wants.

The most interesting unanswered question is what happens to the agents Walmart already works with. Muse, ChatGPT and Gemini are named in the announcement as existing conversational commerce partners, alongside Walmart's own Sparky assistant. Whether those integrations migrate onto the gateway, keep their bespoke deals, or get terms the smaller developers will not get is the tell for how open this really is. A gateway that approves the three biggest assistants and a handful of partners is a distribution deal with a standards document attached.

And the liability question is only half answered. Walmart's terms tell agents what they may not do; they do not say who pays when an agent does it anyway and the shopper did not notice for a week. The returns process covers the product, not the principle. If a misbehaving agent buys a hundred dollars of groceries and the customer only spots it after the food is eaten, the transaction stands, the profile gets deleted, and the argument about whose agent it was has no forum.

What to watch next

Whether the approved-agent list is public. If Walmart publishes it, agentic commerce has a registry and every other retailer has a template. If it does not, "approved" means "whoever we signed," and the terms are decoration.

Whether Amazon answers with a gateway. Amazon's position has been that sites choose who enters, and it has enforced that with blocks and, before that, a lawsuit against Perplexity it lost at the injunction stage. Walmart just showed a version of the same veto that also captures the transaction, the data and the customer relationship. If Amazon copies it, the fight over agent access stops being a fight and becomes a procurement standard.

Whether anyone puts a number on agent error rates. Walmart will learn them first. The moment a retailer discloses the share of agent orders that get returned, the pricing conversation changes — because that number is what an agent channel actually costs, and it is the only number in this story that nobody has published.

If you would let an agent buy your groceries under a hundred-dollar cap, what would you still want to approve by hand? Write to us via our contact page.