Walmart has published the terms on which it will let third-party AI agents buy things on behalf of its customers, and it is the first large retailer to write the rules rather than block the traffic. Nothing in it is open to the public. The gateway that makes it work is an early beta, and Walmart’s own product page states that access is limited, available by invitation only, and may be suspended at any time.
That combination is the story. The rules are real and they are worth reading now, because they are the template the rest of the market will argue about. The thing itself is a closed trial.
What Walmart shipped, and what it only announced
Two pieces were described, and they are at different stages.
Agent Profiles is a separate profile for each AI agent inside the customer’s own account, so that an agent’s purchases are visible and attributable rather than mixed in with the person’s. Walmart says customers will be able to authorise each agent with a dedicated profile. It is not live. The company’s wording is that Agent Profiles are launching in the next month.
The Walmart Agent Gateway is the developer-facing half, built on MCP, the open protocol agents use to connect to outside services. Walmart’s announcement says it will begin with a beta for companies and developers. Its own gateway page is blunter: an early beta program, access limited, invitation only, suspendable at any time.
So one component is a month away by the company’s own account, and the other is a closed programme you cannot join by choosing to. A reader who takes this as a surface they can sell into this quarter has misread it.
Tool comparison · FastMoss vs Kalodata
Agentic buying rewards sellers whose product data is clean enough for software to choose it without a human looking at the page. FastMoss and Kalodata show how products are actually found and ranked on TikTok Shop, which is the same discipline one platform generation earlier.
Comparison coming soon
FTC disclosure: our tool comparisons carry affiliate links. If you sign up through one we may earn a commission, at no extra cost to you. We rank on merit, never on commission, and the verdict is written by a human.
From the publisher
Everything we know about running Amazon, in one course.
The commercial strategy, taught by operators who run our own brands.
The spending limit, and what the dollar figures are not
Walmart describes two modes. A customer either approves every purchase individually, or grants the agent permission to buy within a spending limit they set. The announcement gives examples of that limit: 50, 100 or 200 dollars.
Those three figures have already started travelling as though they were the ceiling. They are not. They are illustrations of a limit the customer chooses, and Walmart publishes no maximum, no minimum and no default. Any sentence of the form “Walmart caps agent spending at 200 dollars” is wrong, and it is wrong in the specific way that matters: it converts a customer-set variable into a platform rule.
The difference is not pedantry for anyone modelling basket sizes on this surface. A customer-set limit with no published distribution tells you nothing about what agents will actually be authorised to spend, and the examples tell you only what Walmart thought sounded reasonable in a press release.
Why the payment-card line matters
The clearest commitment in the announcement is the one about credentials. Walmart states that agents do not receive customers’ payment-card information and cannot add or change a customer’s payment methods.
That is a meaningful boundary and it is worth understanding why it is there. The obvious failure mode for agentic buying is not an agent making a bad purchase, which a spending limit contains. It is an agent holding reusable credentials, which turns a compromised or badly behaved agent into an open line of credit. By keeping card details on Walmart’s side of the boundary and refusing agents the ability to modify payment methods, the design limits the blast radius to transactions rather than to the account.
For a marketplace seller the read-through is structural rather than immediate. If the buying agent never holds the credential, then the agent is a selection and comparison layer sitting on top of the retailer’s checkout, not a parallel checkout of its own. The retailer keeps the payment relationship, and the competition moves entirely to how a product looks to a machine deciding between options.
Being first is the point
There were three options available to a large retailer facing agent traffic, and Walmart has taken the one that is hardest to reverse.
The first is to block, which is what the early posture across retail mostly was: treat an agent as unwanted automation and refuse it at the door. The second is to ignore, letting agents scrape and buy through ordinary consumer flows without acknowledging them, which is where a good deal of traffic quietly sits today. The third is to admit them on published terms, which is what this is.
Publishing terms first has a specific advantage. Whoever writes the first credible rulebook sets the vocabulary everyone else has to argue in, and the vocabulary here is Walmart’s: agents have identities, customers authorise them individually, limits are set by the customer, and credentials stay with the retailer. A competitor that later wants a different arrangement now has to explain why, to a market that has already seen a reasonable-looking alternative.
Tool comparison · FastMoss vs Kalodata
Before a machine can choose your product it has to find it, and that is a category and demand question first. Our head-to-head of FastMoss and Kalodata looks at how each one reads demand and competition across a category.
Comparison coming soon
FTC disclosure: our tool comparisons carry affiliate links. If you sign up through one we may earn a commission, at no extra cost to you. We rank on merit, never on commission, and the verdict is written by a human.
What changes about a listing when the reader is not a person
This is the part that outlives the beta, and it is the reason to read the announcement at all.
A human shopper on a retail site is influenced by photography, by brand familiarity, by review sentiment, by the layout of the page. An agent working to a brief and a spending limit is not. It reads structured attributes, price, availability and delivery promise, and it compares them across candidates in a way no human has the patience for.
What that rewards is dull and specific. Complete and correct attributes. Consistent units and pack sizes. Accurate stock and dispatch data. A price that is competitive on the total landed figure rather than on the headline. What it stops rewarding is the part of merchandising that works by catching an eye.
None of that is new advice, and that is rather the point. Feed quality has been the unglamorous item on every marketplace checklist for years. Agentic buying does not introduce a new requirement, it removes the slack that let sellers get away with ignoring the existing one.
What a seller does differently on Monday morning
Honestly, for a European seller, nothing operational. This is a US retailer’s invitation-only beta, Walmart has not said where it will open or when, and there is no mechanism here that a seller outside the programme can touch.
Two things are still worth doing, and neither depends on Walmart.
- Audit your structured attributes on the channels you already sell on, against the question of whether a machine could choose your product without seeing the page. Missing dimensions, inconsistent pack quantities and vague delivery promises are the things that lose a comparison conducted by software, and they are already losing you some of them today on ordinary marketplace search.
- Read the rules themselves rather than the coverage of them. Agent identity, per-agent authorisation, a customer-set spending limit and a hard boundary on payment credentials are four design decisions that any competitor writing its own agent policy now has to either copy or justify departing from. Knowing what the first published template says is cheap, and it is the only durable thing on offer this week.
Where the honest uncertainty sits
Walmart has not stated which countries or marketplaces this applies to, so the geographic scope is unknown rather than confirmed as US only.
No eligibility criteria for the gateway beta are published beyond invitation, and no enrolment or end dates are given.
Agent Profiles are described as launching in the next month, with no date, and nothing about them can be observed on the public site today.
No volume, pricing or commercial terms for agent-originated transactions have been published, so the economics of the channel cannot be estimated from anything in the announcement.
And Walmart’s own page expands MCP incorrectly, which is a small thing, but it is a reminder that this is a fast-written announcement about an early programme rather than a settled specification.
The three-line version
- If you run the marketplace listings, nothing to do on Walmart and one thing to do everywhere else: check whether your structured attributes are complete enough for a machine to choose you without seeing the page. That is the capability this announcement rewards, and it is already being tested by ordinary search.
- If you have just taken on channel strategy, the gap worth naming is that nobody on the team has read the first published agent rulebook. Agent identity, customer-set limits and the payment-credential boundary are the decisions every competing policy will be measured against.
- If you sign the budget, there is nothing to approve. Refuse any proposal that treats this as an available sales channel, because it is an invitation-only beta with no published geography, no dates and no commercial terms.
Sources
- Walmart Corporate, Making It Possible for Agents to Shop on Behalf of Customers: https://corporate.walmart.com/news/2026/10/Making-It-Possible-for-Agents-to-Shop-on-Behalf-of-Customers
- Walmart Agent Gateway: https://www.walmart.com/agents