The most important platform shifts rarely arrive with a keynote. They arrive as a quiet infrastructure decision that changes who gets to use your store, and on what terms. That is what happened on Wednesday, October 21, 2026, when Walmart announced it would let AI shopping agents buy on behalf of its customers — in the same week Amazon was blocking those very same agents from its storefront.
This is not about one retailer liking robots more than another. It is about two different bets on where shopping is going. If you build websites, run an online store, or write code that touches e-commerce, Walmart's bet affects you directly. The question is no longer whether AI agents will shop for your customers. It is whether your store will be ready when they arrive with a wallet and a shopping list.
What Walmart actually announced
CEO John Furner announced three pieces at the Masters of Scale Summit in San Francisco. Keep them separate, because they do different jobs.
First, the Walmart Agent Gateway: a server built on the Model Context Protocol (MCP), the open standard Anthropic originally created for connecting AI agents to outside tools and data. It gives approved agents a direct way to search Walmart's inventory and place orders, instead of scraping the site or pretending to be a human. The Gateway is starting as a beta for companies and developers, and it covers Walmart and Sam's Club.
Second, Agent Profiles, due over the coming weeks, starting next month. Each agent you use gets its own profile inside your Walmart account, with its purchases shown separately from your own. A profile is created the first time you use an agent through the Gateway, it can be deleted at any time, and it works for any account holder. No Walmart+ membership is required.
Third, published terms for agents on Walmart's site: the rules that decide which agents stay and which get thrown out. Walmart already had integrations with ChatGPT, Gemini, and Meta's Muse, plus its own assistant, Sparky. The Gateway formalises that strategy into one front door with rules attached.
Why the Agent Gateway matters
If you have ever automated anything against a major retail site, you know why a sanctioned gateway matters. Scraping is fragile: a redesign breaks your parser, bot detection blocks your traffic, and half your engineering time goes on pretending not to be a bot. An official MCP server flips that. The retailer says: here is the interface, here are the rules, build against this and we will not treat you as an intruder.
An MCP server also lets the retailer see exactly which agent is doing what. That visibility is what makes everything else possible: spending limits, approval flows, and blocking bad actors all work once agents come through a front door instead of a side window.
One design detail deserves attention. Walmart says the Gateway lets an agent find a customer's usual items, like their regular milk brand, without handing over the customer's wider purchase history or personal data. The hardest problem in agentic commerce is not buying the right thing. It is deciding how much of your life the agent needs to see to buy it. Walmart's answer — as little as possible, scoped to the task — is a template other retailers should copy.
Agent Profiles: the authorized-user idea, applied to bots
Agent Profiles solve a problem I have felt personally. Let a family member use your account and the mess follows: recommendations get polluted, order histories blur, and sooner or later someone asks why you bought something you have never heard of. Replace the family member with an agent that places orders on its own, and the mess gets worse — and more expensive.
Walmart's comparison is apt: a profile works like adding a person as an authorized user on a credit card. You can see what the agent bought, separately from your own orders. You can require approval for every purchase, or set a spending limit instead. Reporting has cited example limits around fifty, one hundred, or two hundred dollars: a weekly-groceries control, not a blank cheque.
That control is what makes this usable for ordinary shoppers. Without it, letting an agent shop for you means trusting it completely, and most people will not do that. With a limit and an approval option, trust becomes graduated. Start with approval for everything, watch the agent for a few weeks, then loosen the leash. That is how new technology gets adopted: not by demanding a leap of faith, but by giving people a dial they can turn slowly.
The rules: what agents can and cannot do
Walmart's terms are short, and that is a strength. Agents must not act beyond what the shopper intends. They cannot access payment details or change a customer's payment method. They may only access the data a task actually needs. Rule-breakers get blocked. And the standard return policy applies to agent orders, so a wrong purchase is not yours to keep.
Read that as a developer, because it is a specification for an agent Walmart will tolerate. Your agent needs a clear statement of user intent, and must stay inside it. It must never touch payment directly. It must practise data minimisation: ask for the milk brand, not the purchase history. And a bad action must have a real consequence — removal, not a warning email.
Vague principles like "be helpful and safe" cannot be built against. A rule that says "you never see the card number" can. If you are building an agent today, Walmart has handed you a checklist worth stealing, even for stores that have not published their own rules yet.
Why Amazon went the other way
Amazon's position is the mirror image. In recent weeks it has blocked outside agents from buying on its store, including Meta's Muse and Perplexity's shopping agent, with reporting also mentioning Google's agents. Walmart welcomes agents on its own terms; Amazon keeps them out entirely. Same technology, opposite answers.
Reporting points to different business models. Amazon earns heavily from sponsored listings and advertising — the paid placements shoppers see while browsing — and promotes its own assistant inside its store. An outside agent that searches, compares, and buys without showing those sponsored pages weakens both the advertising loop and the data link between Amazon and the shopper. Blocking agents protects that model.
Walmart has a smaller share of online sales, so meeting shoppers inside ChatGPT, Gemini, or Muse offers reach it might not win on its own site. Its core business is moving physical inventory: a sale through an agent is still a sale, and the sponsored-listing trade-off is less central. Neither choice is generous or villainous; both are rational given different businesses. That is why this split matters. It is not a disagreement about technology. It is a disagreement about business models, and those are the disagreements that reshape industries.
What this means if you build or run a store
If you run a WooCommerce store or a WordPress site with a shop, you do not need to rebuild anything this week. But start thinking in Walmart's terms, because your customers' agents will arrive carrying those expectations.
First, clean product data matters more than ever. An agent does not browse lovingly. It reads titles, prices, variants, availability, and shipping terms. Messy data — vague titles, missing variants, prices hidden behind three clicks — means an agent will skip you or get you wrong. The stores that win in agentic commerce will have machine-readable catalogues, not just human-pretty ones.
Second, write your own agent rules before someone else's agent writes them for you. One page is enough: which agents you welcome, what they may buy, whether orders need approval, how returns work. Walmart's template — approval or spending limit, no payment access, returns as normal — is a fine starting point.
Third, think about separation. Can your store tell an agent's orders apart from the customer's own? Can the customer revoke the agent's access without deleting their account? You do not need Walmart's infrastructure to answer, but you need an answer, because the first bulk wrong order will prompt exactly these questions.
Finally, do not bet everything on one retailer's rules. Build against open standards where you can — MCP is the obvious one — and keep product data portable. And if you are building an agent, build the consent flow first: approvals, limits, and a clear purchase record. That control is not a compliance chore. It is the product.
What could still go wrong
Do not oversell this. The Gateway is in beta and Agent Profiles have not shipped yet. A control in a press release is not the same as a control that works at midnight when an agent has ordered the wrong thing. The real test is the first messy month: how fast Walmart blocks a misbehaving agent, how clearly it explains the block, and how painless the return is.
Watch who gets approved, too. If approval is slow, opaque, or limited to a few big partners, the open front door starts to look like a VIP entrance. And watch the rest of the industry: will Target, Costco, or others follow Walmart's MCP route, build proprietary APIs, or join Amazon's blockade? A fragmented agent web is the worst outcome for small stores, because only the biggest players can afford a dozen bespoke integrations.
Conclusion
On October 21, 2026, one of the world's largest retailers decided AI agents are customers, not intruders, and built them a front door with rules attached. Amazon decided the opposite. Between those positions, the future of online shopping is being negotiated in infrastructure — MCP servers, agent profiles, spending limits, and published terms — not slogans.
For shoppers, the news is simple: soon you may hand your grocery list to an agent and trust it to stay inside limits you set. For developers and store owners, the brief is clear: clean product data, written agent rules, consent built in first, and open standards so you survive whichever way the industry jumps. Walmart has written its rules. Read them carefully — and get to work.
FAQ
What did Walmart announce on October 21, 2026? An Agent Gateway, Agent Profiles, and published terms for AI agents. The Gateway is an MCP server letting approved agents search inventory and buy at Walmart and Sam's Club; Profiles give each agent a separate account profile; the terms set spending controls and bar agents from payment details.
What is the Walmart Agent Gateway? A server built on the Model Context Protocol that gives approved agents a direct way to search inventory and place orders. It is starting as a beta for companies and developers, instead of agents scraping the site or mimicking human browsing.
When do Agent Profiles launch, and do I need Walmart+? They are due over the coming weeks, starting next month. They work for any Walmart account holder — no Walmart+ needed — and a profile is created on first use through the Gateway and can be deleted at any time.
How do spending limits and approvals work? Shoppers can require approval for every purchase or set a spending limit; reporting cites examples around fifty, one hundred, or two hundred dollars. Agent purchases show separately from your own, and standard returns apply.
Why is Amazon blocking agents while Walmart welcomes them? Reporting points to business models: Amazon earns heavily from sponsored listings and its own in-store assistant, so an outside agent bypassing its pages weakens that loop. Walmart, with a smaller online share, gains reach by meeting shoppers inside ChatGPT, Gemini, or Muse.
What should a small store owner or developer do now? Clean up structured product data, write a one-page agent policy on approvals, limits, and returns, ensure agent orders can be reviewed and revoked separately, and build against open standards like MCP. If building an agent, build the consent flow first.
