On October 21, 2026, Walmart launched Agent Gateway, an MCP server that lets approved AI agents search Walmart and Sam's Club inventory and complete purchases on a customer's behalf. TechCrunch reported the launch under a headline that captures the moment: "While Amazon blocks AI agents, Walmart rolls out the welcome mat." With one move, the largest US retailer settled a question most businesses are still debating. The issue in 2026 is not whether AI shopping agents exist; it is which ones you let through the door, and on what terms.
This article explains what the Agent Gateway actually is, how an agent purchase flows, the controls arriving in November, the conduct rules Walmart is imposing, and what smaller merchants should take from the biggest retailer's decision to welcome agents instead of blocking them.
What the Agent Gateway actually is
The Agent Gateway is a Model Context Protocol (MCP) server. MCP, the open standard first released by Anthropic in late 2024, has become the closest thing the industry has to a universal adapter between an AI application and an outside system. Walmart's server exposes a specific pair of capabilities to agents a customer has approved: searching the retailer's online catalog, and purchasing items for the customer.
Several boundaries announced with the launch deserve precision:
- The gateway works across both Walmart and Sam's Club, the company said in its October 21 announcement.
- Any Walmart account can use it. A Walmart+ membership is not required, which removes the obvious premium gate a company could have used to slow rollout.
- Walmart says agents can surface a shopper's usual brands and items without the customer handing personal information or purchase history directly to the agent. That is a claim about how the connection routes data rather than a verified audit finding, and it deserves independent scrutiny once researchers test the gateway in the wild.
How an agent purchase flows, step by step
- The customer connects an approved agent through Agent Gateway using their Walmart account.
- The agent queries the catalog through the gateway's MCP tools and proposes items or alternatives.
- The customer confirms the selection.
- The purchase is completed on the customer's behalf through Walmart's own checkout rails. Under Walmart's agent conduct rules, the agent never touches payment information.
Walmart positions the privacy angle carefully: the gateway is described as letting an agent find your usual items without you surrendering your purchase history to the assistant's maker. If that separation holds up under testing, it is exactly what security reviewers have been asking for in agentic commerce. If it does not, expect the discrepancy to surface in the first third-party teardown.
The November controls: Agent Profiles
The part security and identity teams will care about arrives next month. Walmart plans to introduce Agent Profiles in November 2026. The first time a customer uses a given agent through the gateway, a dedicated profile is created inside their Walmart account. The shopper can authorize the profile, manage it, delete it at any time, and choose between two delegation modes:
- Approve every transaction. Each purchase decision stays with the human.
- Authorize a spending limit. The agent may buy freely within a customer-set cap.
Walmart compares the arrangement to adding an authorized user to a credit card: the authorized user's transactions are identified in their own lane, and their spending can be capped. It is a deliberately boring analogy, which is what makes it credible. Per-agent identity inside a mass-market retail account is the same governance pattern we described when the big US banks locked agents out of customer accounts in mid-October, except deployed at the front door rather than the wall. Identity first, spending inside boundaries, revocation at any time.
The conduct rules agents must follow
Walmart is also publishing terms for how agents may behave on its site, and three riders matter most to businesses watching the precedent:
- Agents may not take actions beyond the shopper's intent.
- Agents may not access customer payment information.
- Agents may not obtain information they do not need for the task.
Agents that violate the rules will be blocked from the site. And in a line every early adopter should read twice, Walmart says its existing returns process applies to agent-made purchases, including mistakes. That is not a contractual promise that every unwanted order gets refunded; it is the standard policy with a new actor inside it. Expect the gap between "agent error" and "customer remorse" to become legally interesting in 2027.
The welcome mat versus the bot wall, mapped
| Player | Posture toward AI agents | Mechanism | What it signals |
|---|---|---|---|
| Amazon | Blocking | Countermeasures that kept agents like Meta's Muse from completing purchases, per TechCrunch's October 2026 coverage | Defending checkout control end to end |
| Citi, Bank of America, JPMorgan | Blocking | Formal restrictions keeping agents away from customer accounts, announced mid-October 2026 | Fraud and deposit-flight concerns dominate |
| Shopify (with Meta Canvas) | Opening by protocol | WebMCP checkout and open catalog access for personal agents, September 2026 | Merchant choice, protocol-first |
| Walmart | Welcoming with governance | Agent Gateway MCP server, Agent Profiles and spend controls in November, agent conduct rules | Agents treated as a purchase channel that needs identity rails |
The pattern across the table: the blockers are defending a data-heavy, trust-critical surface (bank accounts, checkout control), while the openers are monetizing a discovery surface (catalog, retail intent). The differentiator is whether the company can charge for the traffic once it arrives.
What smaller merchants should take from this
The demand side has already moved. Retailgentic's October 2026 "Vibe Check" survey, reported by Forkast, found roughly half of US shoppers now use AI somewhere in their shopping journey, and that 11.4 percent of observed path-to-purchase volume ran through autonomous agents. Treat the exact figure as directional rather than gospel; treat the direction as settled. The phrase "agentic commerce optimization" exists precisely because catalog and checkout data increasingly has to be legible to software, not just to people.
Four ordered steps for a smaller store:
- Make product data machine-readable. Clean structured product data, stable availability signals, honest pricing. Agents will trip over the same mess that breaks your SEO.
- Write down your agent policy. What may an agent do on your site: browse, compare, resolve, pay? A short public note beats silence, and it is the same posture Shopify took when it opened its catalog to personal agents.
- Decide your posture deliberately. Blocking everything, welcoming narrowly, or waiting are all defensible positions. Drifting into one by accident is not.
- Build the fraud reflexes before the volume. Agentic purchasing inherits the same blind spots we catalogued in our agentic payments fraud analysis: rules tuned for human pace, velocity checks that misread fast machine traffic.
Not every store needs an MCP server this year. Every store that intends to still be discovered in two years needs to know which of its channels, venues and suppliers are agent-ready, and to make that a conscious decision rather than a surprise.
If you sell online and want a clear-eyed read on whether agentic commerce belongs in your roadmap this quarter or next year, that is exactly the kind of planning conversation we help operators run. Start one at our contact page.
