Walmart gave shopping agents paperwork: dedicated profiles, a dedicated gateway, and six rules

The short version: Walmart is not "welcoming all agents" — it is giving agents paperwork

On October 21, 2026, Walmart announced three things: Agent Profiles, the Walmart Agent Gateway, and a set of terms for AI agents. Taken together, Walmart's position is precise: it does not object to AI shopping for you, but it objects to AI shopping by pretending to be you. Until now, agent shopping mostly travelled the browser route: the agent logged in with your account, clicked pages, filled forms, and checked out like a person, and merchants struggled to tell whether the screen showed you or your agent. Walmart's new scheme gives the agent a formal identity, a dedicated channel, and explicit rules: the agent has its own profile, its own spending limit, and its own behavioural boundaries; mistakes go through Walmart's returns process; and agents that break the rules are blocked outright.

The timing is not accidental. Around the same time, Amazon began blocking outside shopping agents, while Walmart chose the opposite direction: not closing the door, but replacing it with a front door that has a registration desk. For consumers, "let AI buy things for me" gains a formal shape for the first time, similar to adding an authorised user to a card. For developers and merchants, agent shopping moves from "scrape my website" to "use my official channel." This article first explains Walmart's three pieces, then where its disagreement with Amazon really lies, then what it genuinely means for vibe builders and independent merchants, and finally which key questions Walmart has not answered.

Piece one: Agent Profiles — a supplementary card for each agent

The core idea of Agent Profiles is one Walmart itself states with a homely analogy: adding someone as a cardholder on your credit card. You can set a spending limit for the cardholder, see what it spent, and turn it off at any time. In shopping terms: each AI agent gets a dedicated profile inside your Walmart account, its purchases are shown separately from your own, you can require approval for every purchase or give it a spending limit — Walmart's announcement gives $50, $100, and $200 as examples, not fixed tiers — and you can switch any agent's permission off at any time.

This solves the messiest part of agent shopping so far: identity confusion. When an agent used your account, the system saw only you; the agent's mistaken orders, records, and recommendations all blended into your account. When something went wrong, it was hard to tell customer service "this order was not mine, it was my agent's," and hard for service to distinguish. A dedicated profile turns that into a checkable fact: which agent this order belongs to, under which limit it happened, and whether it was approved. Walmart also stresses that customer approval is the default, and that Walmart may require additional approvals even for purchases within an agent's spending limit. In other words, a limit is not an "approval-free pass" but a ceiling; approval authority stays with the customer.

Agent Profiles are expected to launch within a month of the announcement, around November 2026. They work for any Walmart account holder, with no Walmart+ membership required. That detail matters: Walmart is not treating agent shopping as a premium-member add-on but as part of shopping infrastructure. Its commercial logic is plain: if you already ask ChatGPT, Gemini, or Meta's Muse to build lists and compare prices, Walmart would rather the final order also happen at Walmart than somewhere else. Giving the agent a formal identity is the cheapest way to keep that order.

Piece two: Agent Gateway — let agents use the machine's road, not the human's

The Walmart Agent Gateway is a server built on MCP, the Model Context Protocol originally proposed by Anthropic and now a common open standard for agents connecting to outside tools and data. Its job is to give agents a faster, more reliable route than simulating a human clicking through web pages: instead of parsing page layouts, handling pop-ups, and guessing button positions, the agent searches inventory and orders directly through the Gateway. Walmart says the Gateway starts as an invitation-only beta for companies and developers, not an immediate opening to everyone.

Its other key design is "help you find your usuals without handing your full history to the agent." Walmart's example is concrete: if you usually buy a particular brand of milk or a sourdough loaf, the Gateway can help the agent find those usual items without sharing your personal information or broader purchase history. That distinction deserves emphasis: previously, an agent helping you shop often needed your full history to guess your preferences; now Walmart makes "find the usuals" a platform capability, so the agent receives only the information needed for the current task, not your entire data. This matches the terms' rule of "only the information needed for the task," and it is the most product-savvy part of the whole scheme.

For developers, the Gateway matters because it turns agent shopping from a grey area into a documented integration. Previously, letting an agent shop for users on a site usually meant browser automation: the site redesigns, the automation breaks, and nobody owns the failure. Now Walmart offers an official channel with a protocol, a beta, and rules. That does not mean every merchant will follow, but it does provide a replicable template: if you want agents to shop in your product, what you need is not a web page but an agent-readable interface, an account system that distinguishes agent identity, and rules that assign responsibility when things go wrong.

Piece three: the terms — what agents may and may not do

Walmart's terms for AI agents list six core rules, worth reading one by one because they read almost like a draft constitution for agent shopping. First, customers are in control, and Walmart's returns process applies — including when an agent does not act as intended. Second, customers must approve agent purchases by default, and Walmart may require additional approvals even within spending limits. Third, agents must not take actions beyond what the shopper intends. Fourth, agents do not receive customers' payment-card information and cannot add or change payment methods. Fifth, agents may access only the information needed for the task and allowed by the customer. Sixth, agents that do not follow the terms will be blocked from Walmart's site.

The easiest to underestimate are the fourth and sixth rules. Not handing payment-card details to the agent means the agent's "purchase" is technically not swiping your card directly but issuing an instruction through Walmart's platform, which completes payment; the agent never touches the money. That differs fundamentally from the older pattern of an agent checking out with your login and payment details, where a failure left it unclear whether the agent, the merchant, or you were at fault. And "violators get blocked" supplies an enforcement mechanism: rules without enforcement are just a blog post; Walmart's explicit statement that it will block non-compliant agents gives the first five rules teeth.

One more clause deserves separate attention: Walmart says the agent's privacy policy applies to information the agent receives, while Walmart handles information the agent provides under Walmart's own privacy notice. Responsibility is split by segment: what you hand to the agent is the agent provider's responsibility; what the agent hands to Walmart is Walmart's. That honesty matters. It reminds users that even the safest Walmart channel cannot substitute for choosing a trustworthy agent: the agent you pick may still collect, retain, or use your data. Choosing an agent and choosing a channel are two different decisions and cannot replace each other.

Walmart also states a pricing principle: prices an AI agent sees are the same prices a human would see on Walmart.com, with applicable offers and fees shown before purchase. The sentence answers a real worry: will there be an "agent price" and a "human price"? Walmart's answer is no — it prices products, not people. That promise currently has no third-party audit mechanism. Writing it into the terms is one thing; enforcing it over time is another, and it is worth tracking in follow-up coverage.

Where Walmart and Amazon actually disagree

Placed side by side, the disagreement is not "one company likes AI and the other hates it." Amazon's stated reasons for blocking outside shopping agents, according to public reporting, are that those agents did not identify themselves, captured user logins, and bypassed Amazon's shopping pages and advertising system. Walmart's approach answers those reasons one by one: identity is solved by Agent Profiles, login is solved by the Gateway's official channel, and page-bypassing is solved by an MCP interface. In other words, the two companies overlap on the diagnosis but prescribe opposite treatments: Amazon blocks agents that arrive without identity, while Walmart issues identity and lets them through the front door.

Behind the disagreement is a business-model difference. A large part of Amazon's online business depends on users browsing the site, clicking sponsored listings, and seeing ads; an agent that orders directly without viewing pages weakens that chain. Walmart's online share is relatively smaller, so it cares more about "wherever the user is, the order follows": if users already decide inside ChatGPT or Gemini, Walmart would rather the agent bring the order in than insist users first open Walmart's site. This is not a moral judgement about which company is more "open"; it is two different calculations from two different positions. For consumers, the near-term result is that the same agent may walk through the front door at Walmart and be stopped at Amazon, and users who shop at both will need different agent strategies for different stores.

It also needs stating that Walmart's announcement currently covers only Walmart and Sam's Club in the United States. Walmart has not said whether Agent Profiles and the Gateway will extend to Canada, Mexico, Flipkart, or other markets, nor has it stated the Gateway beta's selection criteria, pricing, or general-availability date. Those unknowns are not minor: they decide whether this is a US-market pilot or the start of global agentic-commerce infrastructure. Until answers arrive, the accurate description is "a formal US pilot," not "the arrival of the agent-shopping era."

How this differs from our earlier shopping-agent coverage

This site has published two adjacent pieces: one on Meta's Sierra personal-agent protocol, about the "traffic rules" for agent shopping, and one on Amazon blocking the Grok Bot shopping agent, about platform blocking and liability. This piece lands differently from both: Sierra is a protocol-level proposal, the Grok story is a blocking-level conflict, and Walmart is a "open but rule-bound" working system. It is not a sequel to Sierra, nor a reversal story to the Grok event; it is a third answer, in which a platform actively issues identity, sets limits, and writes terms. Read together, the three pieces form the full picture: someone proposes rules, someone blocks violators, someone tries to turn rules into product. Read alone, any one of them would suggest agent shopping has only one possible ending.

What it genuinely means for vibe builders and independent merchants

First, agent readability is becoming a new "listing standard." Previously, for users to find your product you needed search optimisation and a good detail page; now, for an agent to find it and order it correctly, you need machine-readable product information, prices, stock, return policies, and interface documentation. Walmart's Gateway works because structured product data sits behind it; if your product information exists only in images or support chats, an agent cannot reliably complete a purchase for a user. This does not mean you must build an MCP server tomorrow, but it does mean checking first whether your product data is structured and whether prices and stock can be read by a program. If not, agent-era traffic may pass you by — not because you were blocked, but because you are unreadable.

Second, permission design must graduate from "user account" to "agent profile." Walmart's Agent Profiles offer a pattern vibe projects can copy directly: do not let an agent share a user's full account; give each agent a dedicated profile recording who it is, what it may do, what its limit is, and where its operation records live. Even if your project has no agent feature yet, the pattern is worth thinking through now: once users start accessing your product through external agents, you will sooner or later have to answer "was this done by the user or by the agent?" Designing identity separation up front is far cheaper than digging through logs afterwards.

Third, the choice of approval defaults matters more than it looks. Walmart wrote "approval required by default" into its terms — a product decision and a trust decision. The same question faces vibe projects: should your agent feature default to "ask first, then act" or "act first, then report"? The first is safer but chattier; the second is smoother but riskier. Walmart's answer is tiered: small, routine, reversible actions can run within a limit, while large, unusual, irreversible actions require per-transaction approval. That tiering transfers directly: first split operations by reversibility and amount into two tiers, then set a default approval policy per tier, rather than governing everything with one switch.

Fourth, do not confuse "open" with "unregulated." Walmart's openness is conditional openness: identity, limits, terms, and a blocking mechanism. For independent merchants, that means if you plan to welcome agents, do not just write "agents welcome"; write your agent terms: what agents may access, what they may not, what process applies when something goes wrong, and what happens on violation. Openness without terms usually ends as dispute. Walmart's six rules are simple, but they at least provide a starting point worth copying.

Questions still unanswered

Walmart has not stated the Gateway beta's selection criteria, nor whether it will charge, nor whether it will treat merchants of different sizes differently. It has not explained how it detects violating agents: by behavioural analysis, by reports, or by Gateway technical constraints. It has not stated Agent Profiles' data-retention period, nor what happens to purchase records and recommendation data when a user deletes an agent profile. Nor has it said whether the scheme will land in markets outside the United States. These questions are not nit-picking: they decide whether Walmart's scheme is an open standard other merchants can copy or a closed system belonging to Walmart alone. Until they are answered, the fair assessment is: direction worth endorsing, details still to verify.

Conclusion: the watershed for agent shopping is paperwork, not technology

The most important significance of Walmart's announcement is not that it used MCP, nor that it supports ChatGPT, Gemini, and Muse, but that it reframes agent shopping from a technology problem into a paperwork problem: agents need identity, limits, rules, and traceable records. Technology could let an agent order for you long ago; what was missing was the paperwork. Amazon chose to first block agents without paperwork, while Walmart chose to first complete the paperwork and then let them in. Both choices carry costs, but Walmart's choice at least offers a starting point that can be discussed, copied, and improved. For vibe builders, the thing truly worth copying is not any single Walmart feature but its order: give the agent an identity first, and only then talk about letting it buy things for you.