Walmart Opens Its Site to Customer-Owned AI Shopping Agents
Walmart said it will let customers connect their own AI agents to Walmart.com to shop, with systems to approve and review purchases and to set spending limits. The Verge reported the announcement on October 21, 2026, and the capability is set to roll out the following month.
The connections run through a new feature Walmart calls Agent Profiles. Purchases made by an agent require approval by default, and users can set optional spending budgets, with one example capped at $100. Users can review an agent's purchases separately from their own orders and turn the agent off at any time.
Walmart says the integration is built on an Agent Gateway that connects agents through the Model Context Protocol (MCP), an open standard that gives AI models a structured way to reach outside tools and data. According to Walmart, the gateway prevents agents from accessing a customer's shopping history and personal information. Walmart also says agents will be able to find a customer's usual items, such as a particular sourdough loaf or a specific milk brand, without receiving the wider purchase history behind them.
Walmart's terms for AI shopping agents set behavioral limits. Agents must not take actions beyond what the shopper intends, and they cannot change a customer's payment method. Agents may access only the information needed to complete their task, and Walmart says it will block agents that do not follow these terms. Walmart plans to extend agent shopping to Sam's Club as well.
The move follows a sequence of Walmart announcements. In May 2025, Walmart's corporate newsroom said it was developing purpose-built agentic AI tools for retail-specific tasks, using its own data and large language models (LLMs). At its 2025 Investment Community Meeting, Walmart's prepared remarks said digital agents would operate independently and manage complex tasks. Reuters reported in July 2025 that Walmart planned four "super agents" for customers and employees to replace several fragmented AI tools, and that it aimed for online sales to reach 50% of total sales within five years, driven by AI. Walmart describes Sparky, its generative AI shopping assistant, as integrated into the Walmart app.
Walmart's approach stands in contrast to Amazon's. Amazon has blocked Perplexity's shopping agent and Meta's Muse AI assistant from shopping on its site. Amazon sued Perplexity in November 2025, accusing the startup of covertly accessing private Amazon customer accounts through its Comet browser. On March 10, 2026, a court order blocked Perplexity's AI shopping agent from accessing Amazon, and Amazon won the case. On September 23, 2026, Amazon introduced a new agentic AI service for third-party sellers, a separate product aimed at the seller side of its marketplace.
A related effort is also underway on the identity side. On October 6, 2026, Quartz reported that Meta, Walmart and Stripe had launched an open standard for AI agents, a personal agent protocol designed to let businesses verify whether they are dealing with an AI bot or a human customer.
The broader context here is a divergence in how large retailers treat third-party agents. Amazon has enforced its perimeter through litigation and access blocks. Walmart is building a permissioned channel, in which the agent authenticates through a gateway, operates within a user-defined budget, and is subject to contractual limits on what it may do. The two approaches are not mutually exclusive, and the identity protocol Walmart has co-launched suggests that retailers expect to need some way to tell delegated agents from human sessions regardless of which path they choose.
In my view, the questions that matter most for engineers are operational rather than philosophical. Approval-by-default is a sensible control, but its value depends on how clearly an agent's proposed purchase is presented to the user and how reliably the agent resists manipulated instructions embedded in product pages or search results. Scoping data access through MCP reduces exposure of shopping history, but the boundary between "information needed to complete the task" and everything else will be tested in practice. Walmart's commitment to block non-compliant agents will also depend on how it detects them.
The upside, as I read it, is concrete. A shopper who wants to replenish staples can delegate that chore to an agent with a bounded budget, a per-purchase approval step and an off switch, without handing over a payment method or a full purchase history. If the controls hold up under real use, that is a meaningful change in how routine checkout can work, and it gives retailers a way to take part in agent-driven commerce without giving up the storefront.
