The first purchase an AI agent makes at Walmart will not sit in the same place on the account as the ones a customer makes by hand. That separation is the purpose of a set of tools Walmart introduced on Wednesday, and it is the clearest sign yet that the retailer has chosen to compete for agent traffic rather than fence it out.
John Furner, Walmart’s chief executive, presented the tools at the Masters of Scale Summit in San Francisco. The centerpiece is the Walmart Agent Gateway, an MCP server now in beta for companies and developers, which lets approved AI agents search Walmart’s online inventory and complete purchases on behalf of the customer who sent them. Sam’s Club is included from the start.
MCP, the Model Context Protocol, was created by Anthropic as a common way for models to reach outside data, tools and workflows. For Walmart the protocol replaces what agents do today, which is load a page built for people and click through it. That route is slow and error-prone, the company said, and agents get stuck along the way.
The second piece arrives next month. Agent Profiles give each agent its own identity inside a customer’s Walmart account, created automatically the first time that agent shops through the gateway. Purchases land under the profile rather than mixed with the household’s own orders, and the customer sets the terms: approval for every transaction, or permission to buy up to a limit the customer picks, such as $50, $100 or $200.
Walmart describes the arrangement as close to adding an authorized user to a credit card. The profile can be deleted at any time, and a Walmart+ membership is not required to use one.
The third piece is a set of terms for agents, and it is the part other retailers are likely to read closely. Agents may not act beyond what the shopper intends. They do not receive payment card numbers and cannot add or change a payment method. They may see only the information a given task requires. Walmart’s returns process applies even when an agent does not behave as intended, which shifts the risk of a bad automated purchase back to the merchant. Agents that break the terms are blocked.
Walmart also said prices will not differ for agents. “We price products, not people,” the company wrote in its announcement, adding that an agent sees the same prices, offers and fees as a shopper on Walmart.com.
There is a quieter feature with commercial weight. The gateway can help an agent reorder a customer’s regular items, such as a particular loaf or brand of milk, without handing the agent the customer’s broader purchase history. That keeps a decade of household data inside Walmart while still letting the agent do the work.
The contrast with Amazon is deliberate. In September, Amazon blocked Meta’s Muse agent from buying on its site, saying Meta had not disclosed that Muse would shop there and that the agent was capturing customer credentials without identifying itself. Amazon runs its own agentic service, Buy for Me, which it has expanded from 65,000 products at launch to more than half a million. Travel and booking sites have taken the same route, shutting agents out rather than negotiating access.
Walmart had already moved in the other direction. It announced a partnership with Muse in September and offers integrations with ChatGPT, Gemini and its own assistant, Sparky. Furner’s presentation extended that posture into infrastructure: a sanctioned on-ramp, with paperwork.
The timing reflects where the traffic is going. Adobe measured a 127% year-over-year increase in AI-driven traffic to U.S. retail sites in August and expects roughly 130% growth across the holiday season, within a forecast of $275.1 billion in U.S. online sales, up 6.7%. Shoppers arriving through AI tools added items to their carts at a rate 32% higher than other visitors, Adobe found.
For Walmart the arithmetic is straightforward. Global e-commerce sales rose 23% in the quarter ended July 31, and the company is spending to keep that rate up against a rival that controls roughly 40% of U.S. online retail. If a meaningful share of future shopping begins inside a chatbot, the retailer that answers the chatbot’s query keeps the sale. The retailer that blocks the chatbot loses it before the customer ever sees a product page.
The terms also settle an argument about liability that has slowed agentic commerce. Merchants have worried that an agent which misreads a request will leave them holding returns and chargebacks. Walmart’s decision to honor its returns policy in those cases removes the most common objection, at the cost of absorbing some losses.
What Walmart has not done is give agents a discount or a separate catalog. Price parity matters because agents compare. A retailer that quotes an agent a different price than a browser would find itself in a fight it cannot win, and Walmart’s ad business, which sells placement on search results, depends on those results being the same for everyone.
Smaller merchants on the marketplace will now be reachable by agents without negotiating anything themselves, which is both an advantage and an exposure. Their listings will be read by software that sees attributes and price, not photographs.
Analysts said the model resembles the early days of mobile commerce, when retailers built apps for fear of losing customers to whoever owned the phone. The difference is that an agent acts without being watched, and its errors arrive as delivered packages.
Walmart’s answer is to keep the customer’s account as the checkpoint, with spending caps and a returns policy as the backstop. Whether shoppers trust an agent with a card, even one walled off from the card number, is the question the beta is meant to answer.
