Sources checked:
What changed
Walmart said on 21 October that it is introducing Agent Profiles, a Walmart Agent Gateway and terms for AI agents. The company says Agent Profiles are planned to launch within a month and will let a customer authorize each agent separately, distinguish agent purchases from the rest of the account, require approval for every purchase or set a spending limit, and revoke the profile. Walmart also says its planned gateway will give agents a dedicated connection instead of making them use a human web flow. Its published terms say customer approval is the default, agents cannot add or change payment methods, access should be limited to the data needed for the task, ordinary returns apply and noncompliant agents may be blocked. These are Walmart's announced design and launch intentions. The publication does not provide independent adoption, error-rate, conversion, fraud, return or customer-satisfaction results, and it does not establish that every feature is generally available today.
Business implications · analysis
The commercially important change is not that an agent can click a buy button. It is that the retailer is defining a separate transaction actor. That turns agentic shopping from a browser-automation problem into an operating-control problem. A useful control model has at least seven linked records: the customer's identity, the agent's identity, the task and data scope, the approval rule, the spending ceiling, the order audit trail and the revocation or exception path. Returns, substitutions, unavailable items and fee changes must remain attributable after the purchase. Retailers and marketplaces should therefore test the complete transaction lifecycle, not only product discovery and checkout. Merchants supplying such channels should ask which inventory promise the agent sees, when price and delivery are revalidated, who owns an unintended order, and how agent-originated demand is separated from human demand in forecasting and service metrics.
What to watch next
Watch the announced launch rather than treating it as complete. Evidence should include actual profile availability, the beta gateway's documented interface and access process, default approval behavior, revocation latency, exception and return handling, uptime and error rates, and whether agents from different providers behave consistently. Three outcomes are plausible. Dedicated identity and scoped access may make agent orders easier to audit than shared-account automation. Customers may still reject the added setup or approval friction. Or a retailer-specific control model may work well but remain difficult to carry across marketplaces. The useful success measure is not the number of agents connected. It is the share of intended orders completed with correct product, price, delivery promise and authorization, while unintended orders, privacy exposure and unresolved exceptions stay visible.
Primary sources
Company statements are attributed to the publisher. Our analysis is separate; forward-looking points are questions to monitor, not promised outcomes.