Walmart builds a front door for AI shopping agents while Amazon builds a wall
Walmart spent last week doing something no other major retailer has committed to at this scale: building infrastructure specifically so outside AI agents can shop on its behalf, rather than keeping them out. On October 21, the company introduced Agent Gateway, a Model Context Protocol server that lets approved AI assistants — Walmart has named ChatGPT and Gemini as partners — search its catalog and complete purchases for a customer, alongside a separate shopping partnership with Meta's Muse. Within about a month, Walmart plans to add Agent Profiles, giving each AI assistant a dedicated sub-account inside a customer's Walmart login so its purchases are tracked separately from the customer's own.
The timing is pointed. Amazon has spent the same stretch doing the opposite: blocking Muse from operating on its site, suing Perplexity last November over its Comet browser agent, and steering customers toward its own Alexa for Shopping and "Buy for Me" tools instead. Walmart's announcement, and the coverage of it, explicitly frames the company as taking the "welcome mat" approach where Amazon is building a wall. That makes this less a product launch than a live experiment in two incompatible enterprise strategies for the same emerging channel.
Governance as the product, not an afterthought
What's notable about Walmart's approach is how much of the announcement is about constraints rather than capability. Agent Profiles let a customer require approval on every agent purchase, or instead set a spending ceiling, such as $50, $100 or $200, that an agent can spend autonomously, a setup Walmart itself compares to adding an authorized user to a credit card. Walmart's terms for participating agents bar them from accessing a customer's payment-card information, require agents to stay within the shopper's stated intent, and threaten to block agents that don't comply. The existing returns process still applies even when an agent's purchase doesn't match what the customer wanted.
That is a governance framework wrapped around a commerce feature, built proactively rather than bolted on after a problem surfaces. It's an interesting contrast with how agentic AI adoption is unfolding inside most enterprises right now, where governance tends to lag deployment rather than precede it. Walmart is effectively betting that the retailer who sets the rules for how agents can transact, rather than the one who keeps them out entirely, ends up controlling more of the resulting revenue and the risk around it.
A crowded week for purchasing agents
Walmart's move didn't happen in isolation. The same week, Anthropic previewed Claude Duet, an agent that runs in the background across a user's connected apps and can make online purchases with approval, a feature aimed at the same category of task Walmart is now building rails for. Meta's Muse, which launched in September in the US and Canada, has reportedly passed 10 million downloads according to third-party tracking cited in coverage of the announcements, and OpenAI rolled out its own always-on agents, branded Dots, the month before. None of this means Walmart and Anthropic are integrated, but it underscores that the shopping-agent category now has several well-funded entrants simultaneously pushing retailers to decide, quickly, whether to open their doors or lock them.
That decision isn't trivial for a retailer at Walmart's scale. Letting third-party agents complete purchases means ceding some control over the customer interaction, the thing retailers have spent two decades optimizing through search bars, recommendation engines and checkout flows. Walmart's own executives have talked about agentic, multimedia shopping eventually displacing the search-bar-and-results-list model that has defined e-commerce since its inception, which suggests the company sees this less as a defensive move and more as a bet on where discovery is headed.
Why this is an enterprise-adoption story, not just a retail one
It's tempting to file this under consumer shopping news, but the mechanics are pure enterprise AI governance: Walmart had to design an identity and permissioning layer for non-human agents, decide what data an external AI system is and isn't allowed to touch, build spending controls comparable to corporate card limits, and create an enforcement mechanism for agents that violate the terms. That's the same set of problems any large company faces when it lets AI agents act on sensitive systems internally, just applied to a storefront instead of an ERP system. Walmart's public, detailed answer to those questions, offered before any high-profile agent-shopping incident has forced the issue, is a useful reference point for how one of the largest enterprises in the world is choosing to operationalize agentic AI risk rather than simply hoping for the best.
The open question is whether Walmart's bet pays off before the risks it's trying to manage materialize. An agent that misreads intent, overspends past a limit due to a bug, or gets compromised to exfiltrate purchase history is still a plausible failure mode even inside a permissioned system. Walmart is wagering that building the guardrails now, while the agent-shopping market is still small, is cheaper than retrofitting them later. Amazon is wagering the opposite: that the safest posture is to not let the problem in the door at all. Both companies will have real data on who was right well before most enterprises have to make the same call about agents touching their own internal systems.