Walmart will let AI agents shop on its site for customers, and it is building the plumbing to make that safe. On October 21, 2026, the retailer announced Agent Gateway, a beta Model Context Protocol (MCP) server for approved agents, and Agent Profiles, a per-agent account area with approval prompts and spending limits that is due next month. CEO John Furner announced the technology at the Masters of Scale Summit, according to TechCrunch, and The Verge covered the details.
The strategic contrast is the story. Amazon has been closing its doors to outside shopping agents, as we covered when it blocked Meta's Muse shopping agent. Walmart is opening a managed door: agents are welcome if they follow written rules, and customers get a dashboard to supervise them. Below: what was announced, how the controls work, how it stacks up against other retailers, and what builders of shopping agents should do.
TL;DR: Walmart's agent commerce setup
| Question | Answer |
|---|---|
| What is it? | Agent Gateway, an MCP server for approved AI agents to search and buy at Walmart and Sam's Club |
| Status | Agent Gateway is in beta now; Agent Profiles roll out next month |
| Who controls spending? | The shopper: approve every purchase, or pre-authorize up to a budget such as $100 |
| Can agents see my history? | Walmart says it surfaces usual items without sharing broader shopping history |
| What are the rules? | No actions beyond the shopper's intent, no changing payment methods, only the data a task needs |
| What if agents break them? | They are blocked from Walmart's site |
| Do I need Walmart+? | No, any account holder can use it |
What Walmart announced
Per the reports, Agent Gateway connects AI agents to Walmart through MCP, the open standard Anthropic created for plugging agents into tools and data. If you want the background, our MCP explainer covers how servers, tools and clients fit together. Walmart says an MCP route lets agents shop without the hiccups they hit when driving a browser, and keeps them away from your shopping history and personal data.
The company said it already offers integrations with ChatGPT, Gemini and other popular AI apps, and it previously announced a partnership with Meta's Muse agent. The same shopping rails extend to Sam's Club.
Agent Profiles: the part shoppers will see
Agent Profiles put the controls in the customer's hands. A profile is created the first time an agent shops through the Gateway, and the customer can link it, set controls or delete it at any time. By default Walmart requires approval for every agent purchase. Shoppers can instead let an agent buy within a budget, for example up to $100, and review the agent's purchases separately from their own. Walmart compares the model to adding an authorized user to a credit card.
That analogy is useful. A card with an authorized user has a limit, a statement and a way to revoke access, and none of that requires trusting the user's judgment forever. The design separates agent activity from human activity in the audit trail, which is the thing most agent checkout experiments have lacked.
The rules for agents
Walmart published terms that agents must follow. The ones reported: agents "must not take actions beyond what the shopper intends," cannot change a customer's payment method, and may only access the information needed to complete a task. Walmart's normal returns process still applies when an agent does not act as intended, which tells builders that errors by agents will be treated as ordinary customer orders to be returned, not as fraud. Non-compliant agents get blocked.
For anyone building a shopping agent, those terms translate into engineering requirements:
- Intent logging. Keep a record of what the user asked for so you can show an action stayed within it.
- No payment changes. Treat payment method updates as out of scope for the agent entirely.
- Data minimization. Request only the fields a task needs and discard them afterward.
- Graceful approval flows. Design for "pending approval" states, since approval is the default.
- Budget awareness. Handle spend caps as a normal outcome and report them clearly.
Why this matters: two models of agentic commerce
Retailers are splitting into camps. Amazon's approach is exclusion: it has blocked Perplexity's shopping agent and Meta's Muse assistant, and analysts such as Paul Graham have argued that bans create startup openings, as in our post on Amazon banning agents. Walmart is choosing a supervised-access model. Others sit in between:
| Retailer or platform | Approach to outside agents | explainx.ai coverage |
|---|---|---|
| Walmart | MCP gateway, agent profiles, written terms, block violators | This post |
| Amazon | Blocks outside shopping agents | Amazon blocks Meta Muse |
| Shopify | Catalog opened to any agent with Shop rules and budgets | Shopify Shop for any AI agent |
| Universal cart inside AI Mode and Gemini | Google Universal Cart | |
| Stripe | Link payments for agent transactions | Stripe Link agentic transactions |
Walmart's move also lines up with the Personal Agent Protocol standard that Sierra, Meta, Walmart and Stripe backed, which aims for common rules about how personal agents identify themselves and act for people.
Why an MCP server beats a browser agent for shopping
Most shopping agents today work like a person with a very fast mouse: they load product pages, parse layouts, click buttons and hope the checkout flow has not changed. That is brittle in two ways. Pages change, so the agent breaks, and a retailer cannot tell a helpful assistant from a scraper, so it blocks both. An MCP server turns the store into a small set of typed tools, such as search for an item, add it to a cart, and place an order, each with defined inputs and outputs. The retailer decides what each tool returns, which is how Walmart can say the agent never sees the shopper's wider history.
For retailers this is also a bargaining position. A gateway gives them one place to authenticate agents, rate-limit them, log what they did and revoke access, instead of fighting traffic that looks like a bot. For agent builders the cost is that you are now a registered integration partner subject to terms, not an anonymous visitor.
A worked example of the flow
Take a shopper who tells an assistant to reorder the usual groceries for the week. Under the described design, the assistant connects through Agent Gateway and a profile is created the first time it shops. Walmart helps it find the usual brands without handing over the full order history. The assistant assembles the cart. If the shopper kept the default setting, the order waits for an approval tap. If the shopper set a $100 budget, orders under that go through and appear in the agent's own purchase list. This is an illustration of the announced mechanics, not a recorded session, and the real interface may differ when it launches next month.
What people are asking
Does this mean agents can now spend my money freely? No. Approval is the default, and any budget is something the shopper sets, can change and can switch off.
Is Walmart letting every chatbot in? Not exactly. Reports say approved agents and a written code of conduct, with blocking for violations.
Why would Walmart do this when Amazon blocks agents? The reporting frames it as differentiation: if shoppers delegate buying to assistants, being the store those assistants can use easily is an advantage, while a closed store risks being skipped.
Will prices be the same for agents? Not addressed in the coverage, which is one of the open questions below.
What we do not know yet
Several details were not in the reports we could read, and readers should not assume them:
- Which agents are approved. Walmart says "approved" agents, but the selection and review criteria were not described in the coverage.
- How blocking works. Walmart will block non-compliant agents, but how it detects them and how appeals work is unstated.
- Pricing and ads. Whether agent-driven shopping changes sponsored placement, substitutions or promotions is not covered. This matters because a shopper's agent may see a different ranking than a human does.
- Liability for mistakes. Returns apply, but the report does not say who bears cost when an agent buys the wrong thing at scale.
- Beta limits. Rate limits, supported categories and regions for the Gateway beta were not detailed.
There is also no independent measurement yet of how often agents succeed or err on Walmart's catalog. Treat any claims about conversion or reliability as unverified until Walmart or a third party publishes data.
What builders and shoppers should do
If you build agents, apply for access and design to the written terms from day one, because enforcement is blocking, not warnings. Prefer the MCP route over browser automation: Walmart's pitch is that the structured interface is more reliable and exposes less data. Keep your agent's logs, since Agent Profiles will make its purchases visible to the user.
If you shop with an assistant, start with the default of approving every purchase, then raise a budget only for items you reorder, such as groceries. Review the agent profile periodically and delete it when you stop using the assistant. Because agents act on instructions found in pages and tool outputs, remain cautious with any agent that reads third-party content while it holds spending power, a risk class covered in our MCP security explainer.
Related reading
- Amazon blocks Meta Muse shopping agent
- Paul Graham on Amazon banning agents
- Shopify opens its catalog to any AI agent
- Google Universal Cart in AI Mode and Gemini
- Personal Agent Protocol: Sierra, Meta, Walmart, Stripe
- What is MCP? Complete guide
- MCP rug pull and tool poisoning explained
Details are as reported on October 21 and 22, 2026 by TechCrunch and The Verge, and may change when Agent Profiles launch next month.