Walmart opens its doors to AI shopping agents: Agent Gateway with approval and spending limits launches next month, parting ways with blocker Amazon

While Amazon started blocking AI shopping agents, Walmart chose to open the doors wider: on Wednesday the retail giant announced Agent Gateway and other products to manage transactions and companion capabilities for AI agents shopping on its site — starting next month, users will be able to have an AI agent place orders at Walmart, with systems to approve and review purchases and set spending limits^[1][2].

Two routes: blockade vs open door

The Verge's headline says it all: "While Amazon shuts out AI agents, Walmart is letting them in"^[2]. Amazon previously blocked Perplexity's shopping agent and Meta's Muse AI assistant; TechCrunch notes the same lockdown posture at travel and booking sites — facing growing consumer adoption of agents, their answer is to keep AI out^[1]. Walmart runs the other way: it had already struck a partnership with Meta's Muse and offers integrations with popular AI apps like ChatGPT and Gemini, and now adds a layer of agent-era infrastructure^[1]. The real difference is a theory of traffic: the blockers see AI agents as eroding their own entrances and ad slots — when an agent decides for the user, the platform loses display, comparison and upsell opportunities; the open-door camp sees agents as a new transaction channel — as long as checkout happens inside their rails, it doesn't matter who clicks. Walmart's calculus shows through: as the share-chaser in e-commerce, the marginal value of any new order source outweighs the loss of gateway control.

Mechanism design: approvals, limits and "house rules"

Notably, Walmart's openness is conditional. The system lets users approve and review purchases and set spending limits — the user delegates execution, not decision-making, with humans signing off at the key nodes^[2]. And agents must follow the retailer's rules of engagement: orderly access, compliant transactions, auditable action paths — effectively translating the traditional e-commerce user agreement into an interface contract for the agent era^[1][2]. The design rhymes with Claude Duet's approval mechanism, released the same day — "AI acts, humans approve" is becoming the universal safety paradigm of agent commerce: whether Anthropic's agents or Walmart's gateway, the approval chain solves the same problem — giving AI the power to act without losing control.

Commercial significance: the first big test farm for agentic commerce

Set in this week's context, the weight becomes clearer: Meta just disclosed on its earnings call that Muse users average purchases above $200 with first-week volumes so heavy Stripe credit limits had to be raised — transactional agents' demand is proven; the only question left is where it lands. By opening now, Walmart effectively turns itself into North America's largest agentic-commerce lab: if agent-driven conversion and basket-size data look good, the whole industry's stance will loosen; if agent orders come with high return rates or abuse, the blockers' case strengthens sharply. For consumers, the near-term change is simply convenience — hand "monthly household restocking" to an agent, auto-completed within limits; for brands and sellers, the new discipline is agent optimization: when comparison, selection and checkout are all executed by AI, the persuasion target of a product page shifts from humans to models, and structured data, price competitiveness and fulfillment scores become the new ranking currency.

What to watch

First, the actual scope of next month's agent-shopping launch and the first partner-agent roster. Second, Amazon's response — fortifying the blockade or opening selectively under pressure (its own Alexa/Rufus strategy will drive the choice). Third, agent orders' conversion and return data — the first hard read on agentic commerce's quality. Fourth, the fight over "agent identity" standards — who defines the line between legitimate agents and scrapers (an agent.txt debate in the making), with Walmart's rule text a key reference frame^[1][2].

One more constraint can't be ignored: privacy and liability. An agent ordering for a user means consumption preferences, budget boundaries and household inventory data all flow through both the agent and the platform — with no mature rules yet on data ownership, retention or misuse; and when an agent orders wrong (wrong item, over-buying, manipulated), whether liability sits with the user, the agent developer or the platform is equally blank. Walmart's approval and spending-limit mechanisms are a pragmatic first step, but for agentic commerce to go mainstream it needs a full trust infrastructure, akin to e-commerce's early "payment security" narrative.