Amazon Slams the Door, Walmart Rolls Out the Red Carpet: The Battle for Agentic Commerce Is Here

Published on October 22, 2026 — AI & Tech Daily

There is a war being fought right now over the future of online shopping, and the front line is not a price war, a logistics network, or a delivery drone. It is the question of whether artificial intelligence agents — the digital assistants that browse, compare, and buy on our behalf — will be welcomed into the world’s biggest storefronts, or locked out like bots and scrapers. This week, the battle lines snapped into focus with startling clarity. On one side stands Amazon, the undisputed king of e-commerce, which has spent the past month actively blocking Meta’s AI agent Muse from its retail site. On the other side stands Walmart, the world’s largest retailer by revenue, which on Wednesday unveiled a sweeping new technology stack — including an “Agent Gateway,” per-agent spending controls, and a formal terms-of-service for robots — explicitly designed to make AI agents first-class citizens of its stores.

The contrast could not be more dramatic, and the stakes could not be higher. Meta revealed during its third-quarter earnings call that Muse has already crossed 10 million users, that more than a quarter of its engaged shopping users are dropping an average of over $200 per order, and that the company believes the AI agent will eventually generate more revenue per user than Facebook, Instagram, WhatsApp, and Messenger combined. In other words, agentic commerce — the practice of letting AI shop for you — is no longer a science experiment. It is a revenue line, a strategic battleground, and quite possibly the next great platform shift in how the world buys things. And the decisions being made this week in Bentonville, Seattle, and Menlo Park will determine who controls it.

Walmart Opens the Door: The Agent Gateway

Speaking at the Masters of Scale Summit in San Francisco on Wednesday, Walmart president and CEO John Furner announced a suite of new technologies designed to prepare the retailer “for a world where AI agents become mainstream,” according to TechCrunch’s reporting. The centerpiece is the Agent Gateway, a new Model Context Protocol (MCP) server that allows approved AI agents to connect directly to Walmart’s systems.

MCP, originally created by Anthropic, has quickly become the de facto standard for letting AI agents plug into external data sources, tools, and workflows — think of it as a USB-C port for the agentic web. Walmart’s implementation lets an AI agent search across the retailer’s online inventory and make purchases on behalf of its user, all through a sanctioned, structured channel rather than a fragile browser session. Crucially, the retailer says the Agent Gateway can help customers shop for their usual brands or items without having to share their personal information or purchase history directly with the agent. That is a significant privacy design choice: the agent gets the ability to transact, but not a dossier on your shopping habits.

The move extends Walmart’s earlier commitments to the agent ecosystem. The company had already announced a partnership with Meta’s Muse at Meta’s Connect developer conference in September, and it offers integrations with popular AI apps including ChatGPT and Gemini. The Agent Gateway, however, is a more fundamental step: instead of negotiating with individual agents one at a time, Walmart is building the infrastructure to serve all of them. The technology applies across both Walmart and its warehouse club division, Sam’s Club, meaning the agent-first future is being wired into the company’s entire retail footprint.

Agent Profiles: Credit-Card-Style Guardrails for Robot Shopping

One of the most interesting — and arguably most consumer-friendly — pieces of Walmart’s announcement is Agent Profiles, rolling out next month. The feature gives shoppers granular control over what their AI agents can and cannot do, addressing the single biggest psychological barrier to agentic commerce: the fear of a misbehaving bot emptying your bank account.

Here is how it works. The first time you use a given AI agent to shop through the Walmart Agent Gateway, a dedicated Agent Profile is created in your Walmart account. That profile is linked to your account, and the retailer has built technology to distinguish purchases initiated by an agent from those initiated by you directly — so your statement reads clearly, and an agent’s orders can never masquerade as your own.

But the real innovation is the control surface. Customers can choose between two approval modes:

Approval per transaction: every purchase proposed by the agent requires an explicit human confirmation before it goes through.
Spending limits: the customer authorizes the agent to transact up to a set cap, beyond which the agent is automatically stopped.

The retailer explicitly compared the system to adding someone as an authorized user on your credit card: their transactions are identified separately, and you get to cap how much they can spend. Agent Profiles can be deleted at any time, and — notably — agent support is available to anyone with a Walmart account, with no Walmart+ membership required. That last detail matters: Walmart is not gating the agentic future behind a subscription wall, which signals that it views agent support as table stakes for customer acquisition rather than a premium feature.

The Terms of Engagement: What Walmart Demands From Its Robots

Walmart has also introduced formal “terms” for AI agents — a code of conduct for software, which is a genuinely new legal and technical frontier. The terms specify how agents must operate and what rights they have to transact on the site, and they read like a bill of rights for both the shopper and the retailer. Among the key provisions:

Agents cannot take actions beyond what the shopper intends — no scope creep, no surprise add-ons.
Agents cannot access the customer’s payment information — transactions are mediated by Walmart’s own trusted infrastructure.
Agents can only access the information they need for the given task, not the customer’s full account history.
Agents that do not comply with the terms will be blocked from accessing Walmart’s systems.
Walmart’s returns process applies even when an agent doesn’t act as intended — a quietly revolutionary clause that reassures early adopters they will not be stuck with unwanted purchases caused by a software bug.

That last point is the kind of customer-friendly condition that could tip the scales for skeptical consumers. One of the recurring anxieties of agentic commerce is liability: if your AI agent books the wrong item, buys a size too small, or misreads your instruction, who eats the cost? Walmart’s answer is, in effect, “we do.” By extending its standard returns policy to cover agent mishaps, the retailer removes the single biggest financial risk of letting software shop for you — and it is a competitive weapon Amazon, with its famously aggressive stance toward misuse, would find hard to match.

Amazon’s Walled Garden: The Case for Saying No

Against Walmart’s open-arms strategy stands the industry’s 800-pound gorilla. In late September, Amazon began blocking Meta’s Muse AI agent from its retail site, according to Bloomberg’s reporting, meaning the agent could no longer browse or purchase from Amazon’s product catalog. The move was a deliberate policy decision, not a technical glitch, and it sent a clear signal to the industry: Amazon intends to control the agentic layer of its own commerce rather than cede it to third-party assistants.

Amazon’s reasoning is not hard to infer. The company has spent two decades building the most sophisticated shopping experience on earth, with its own AI (from Rufus to its recommendation engines) deeply woven into the customer journey. An external agent interposing itself between Amazon and its customers would capture the relationship, the data, and the commerce economics. Amazon also has a well-documented allergy to automated traffic of any kind, rooted in years of fighting scrapers, price-comparison bots, and fraud networks. For Amazon, an AI agent browsing the catalog looks dangerously like a bot mining the platform — and the platform has every incentive to keep the toll booth on its own property.

But the strategy carries risk. By blocking agents, Amazon is implicitly betting that consumers will keep coming to the site directly — a bet that looks increasingly shaky as the center of gravity in consumer AI shifts from search boxes to proactive assistants. If Muse and its rivals become the default way a generation of shoppers discovers and buys products, a retailer that is invisible to agents is a retailer that is invisible, period. Walmart’s bet is the opposite: ride the agent wave, become the most agent-friendly retailer on earth, and harvest the volume that flows through it. One of these two theses is going to be tested at enormous scale over the next 24 months.

Meta’s Muse Crosses 10 Million Users — and the Carts Are Big

The reason both Amazon and Walmart are fighting over this ground became undeniable on Wednesday, when Meta held its third-quarter earnings call. The numbers Zuckerberg shared with analysts rewrote the conventional wisdom about AI agents being a toy for early adopters. According to TechCrunch’s coverage of the call:

Muse has surpassed 10 million users, confirming estimates released earlier this week by Sensor Tower, which noted that reaching 10 million U.S. downloads in 40 days is a milestone achieved by only 11 other apps in history.
About 25% of people who set up Muse now use it daily, a retention rate that would be the envy of most consumer apps.
Over a quarter of Muse’s engaged users who were shopping via the agent had an average purchase size of more than $200 — a figure that surprised even industry insiders.
Plaid, the financial technology company that connects payment apps to bank accounts and powers Muse’s payment rails, told Meta it “hadn’t seen activity like they’re seeing with their connector on any other service before,” Zuckerberg said.

Zuckerberg declared that Muse has achieved “product-market fit” and that Meta’s focus will now shift to scaling the agent, increasing revenue, and optimizing compute costs. The ambition is staggering in its clarity: Meta is forecasting that Muse will eventually bring in more revenue per user than its family of apps users do today — a figure currently in the “several hundreds of dollars annually” range. The monetization model, as outlined, is a mix of keeping the app free for most users while taking a percentage of the commerce flowing through it, and charging power users for access to the most advanced models and larger token allowances.

On the cost side, Zuckerberg acknowledged that Muse is currently “very compute-intensive to run,” but said Meta is building capacity to serve more than a billion people over the next few years and expects to improve cost efficiency by more than 10x in the next year, which “will enable it to be profitable once the revenue lines are at scale.” He also pointed to the trust infrastructure underpinning the push: the existing “Muse Secure VM” for data protection, and a forthcoming “Muse Confidential VM” designed to prevent even Meta itself from seeing users’ content.

The Blocking Epidemic: CAPTCHAs, Airlines, and the Cloudflare Factor

Amazon is not alone in slamming the door — it is simply the most prominent member of a rapidly growing club. As TechCrunch documented in a deep investigation earlier this month, consumers across social media are reporting that their AI agents are being blocked by a wide range of brands, including Yelp, eBay, Zillow, Pizza Hut, Adidas, and numerous airlines. Some blocks are deliberate policy; others are collateral damage from traditional anti-bot defenses designed for a pre-agent era.

Yelp, for instance, told TechCrunch that it does not permit non-human traffic on its platform unless the agent has paid to access its content through its data licensing program — a position that effectively forces agents to become paying customers. eBay said its approach is not to prohibit all third-party shopping agents, but to restrict unauthorized agents and automated scraping and model training. Zillow said it allows AI agents acting on behalf of consumers, though its anti-scraping measures can occasionally block some. Airlines have been notably hostile territory: Delta said it is “continuing to evaluate” external AI agents while taking steps to protect customers from unauthorized automated activity, while United pointed to terms of use prohibiting automated access without prior written permission.

And then there is the Cloudflare factor. Many users suspect that Cloudflare and other content delivery networks are disrupting agent traffic as a side effect of a September 15 change to crawler defaults that lets site owners block AI training while allowing other bot types. Cloudflare, which now runs a marketplace where AI bots must pay for the data they access — and which recently began testing a browser built specifically for AI agents — occupies a uniquely conflicted position as both gatekeeper and toll collector of the agentic web. The result, for now, is chaos: consumers cannot tell whether a blocked purchase is intentional policy, an anti-bot false positive, or a CAPTCHA that simply nobody (human or machine) could solve in time.

The Quest for an Open Standard: From MCP to a Commerce Protocol

The industry’s answer to this fragmentation is, predictably, a standard. A coalition including Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE, and Decagon is working on an open protocol that would dictate how AI agents communicate with businesses specifically for online commerce — an agent-to-agent layer that separates the good bots acting on a user’s behalf from the bad ones. The effort builds on the same philosophical foundation as Anthropic’s MCP, which Walmart’s Agent Gateway already implements, but pushes further into the transactional domain: authentication, authorization, spending limits, and dispute resolution between agents and merchants.

Meta’s framing of the initiative is worth quoting in full, because it captures the strategic stakes for the entire industry: “Turning away a personal agent means turning away the customer behind it. We’d rather work with businesses to address the underlying concerns than see them lose that customer. There is a path that is good for users and good for businesses. It starts with clear norms, gives businesses more predictable control, and evolves toward a more efficient way for agents to work together.”

Clear norms and predictable control are exactly what Walmart’s Agent Gateway and Agent Profiles deliver in practice. The retailer is effectively implementing, ahead of the standard, the governance layer the whole industry will eventually need: authenticated agents, per-agent profiles, spending caps, transaction approval, and explicit terms with real consequences for noncompliance. When the open standard arrives, Walmart will already be speaking its language — and that first-mover position is precisely the moat the company is trying to build.

Beyond Shopping: The ‘Muse Network’ and the Agent-to-Agent Economy

If the near-term battle is about who gets to serve agents, the longer-term vision is about agents serving each other. Zuckerberg used the earnings call to tease a concept he called the “Muse network” — taking the assistant from, in his words, a “single-player game” to one where “your Muses can work with your friends’ and your families’ Muses.” The implications ripple far beyond retail: shared agents coordinating group purchases, family calendars, joint travel planning, and collective bargaining with businesses.

“And if you want to buy something,” Zuckerberg said, “it’s not just navigating to that business’s website; it’s talking to their agent.” That single sentence is the clearest articulation yet of where the industry is heading: a web where the unit of interaction is not the page visit but the agent conversation. Websites become endpoints; agents become interfaces. In such a world, the retailers that win are not the ones with the best homepage, but the ones whose agents interoperate best with the consumer’s agent — which is to say, the ones that, like Walmart, invest today in the protocols, the profiles, and the trust infrastructure of the machine-to-machine economy. It is also why Meta’s Reality Labs losses — improving from $19.1 billion in 2025 to a projected $15 billion this year — remain strategically acceptable to investors: smart glasses are seen as the natural hardware home for always-on agents, and the agent network is the software flywheel that makes them indispensable.

Meanwhile, AI Builds AI: Recursive’s Flint and the Self-Improvement Race

No daily roundup of this week’s AI news would be complete without the other headline that shook the developer world on Wednesday: Recursive released Flint, a free coding model that runs on a laptop and was — by the company’s own account — built largely by AI. Recursive, the startup founded by Richard Socher (Salesforce’s former chief scientist) that emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from investors including GV, Nvidia, and AMD Ventures, says its automated research system, the “Eureka Machine,” contributed to nearly every step of Flint’s development, from pretraining to post-training to inference.

Flint’s specs are impressive for something that fits on a laptop: it is a mixture-of-experts model with 40 billion parameters, of which only 2.5 billion are active per task — light enough for an ordinary machine and radically smaller than Meta’s 816-billion-parameter Muse Spark Open released last week. In Recursive’s tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple’s M4 Pro chip — roughly double the 58 to 69 tokens per second of comparable models like Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on the same hardware — and topped 150 tokens per second on Nvidia’s RTX Spark. On SWE-Bench Verified, the benchmark for real-world software bug fixing, Flint outperformed every other model under 50 billion parameters. Wired’s Will Knight, who tested Flint on a Mac with 48GB of memory, reported that it nailed his coding requests every time. The Eureka Machine, meanwhile, contributed what Recursive describes as a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels that accelerate underlying computations.

Flint is an early public test of recursive self-improvement — the idea, at once exhilarating and unsettling, that AI can build better AI that builds better AI. Open AI reported last month that its own agents were logging 3.1 workdays for every one worked by a human researcher by mid-August; Recursive’s $410 million compute deal with Amazon Web Services was, in Socher’s words, “less about headcount and more about agent count.” The long-term goal, Socher told Wired, is for the Eureka Machine to achieve full recursive self-improvement with humans overseeing the process and able to stop it if something goes wrong. For now, the field’s cautious answer to “can AI build AI?” is a qualified, benchmark-backed yes — delivered on a laptop you could buy at any electronics store.

What It All Means: The Agentic Web Is Inevitable

Look at the week’s news as a single picture, and the outline of the next era of the internet comes into focus. Consumer AI agents have crossed the adoption chasm: ten million users in forty days, $200 average orders, daily engagement that product managers dream about. The infrastructure is being built in real time: Walmart’s Agent Gateway and Agent Profiles, the Meta-Walmart-Stripe coalition’s commerce protocol, Anthropic’s MCP spreading as the connective tissue of the agentic web. And the strategic responses are polarizing exactly as they did in earlier platform shifts — open versus closed, standards versus walled gardens, partnerships versus blocking.

For consumers, the near-term winner of the Amazon-Walmart divergence is obvious: Walmart’s approach — transparent agent profiles, spending caps, human approval options, and a returns policy that covers robot mistakes — is the consumer-friendly one. Amazon’s approach maximizes control, but it also maximizes friction in a world where the shopper’s loyalty is increasingly to their assistant, not their retailer. History suggests that when a new interface layer emerges (search engines, app stores, social feeds), the platforms that embrace it early tend to capture disproportionate value, while the incumbents that fight it spend years playing catch-up.

The deeper lesson is that agentic commerce is not a feature being added to the existing web; it is a new web being grown on top of it. The retailers, airlines, and platforms that are building the norms, the standards, and the trust infrastructure today — Walmart chief among them — are effectively writing the constitution of that new web. Those, like Amazon, that are building walls will have to hope the constitution favors the walled garden. Given how quickly ten million users adopted a shopping agent, and how much money is already flowing through it, the odds increasingly favor the open door.

This article was written on October 22, 2026, based on reporting from TechCrunch (Sarah Perez, Aditya Mehta, Julie Bort, Ram Iyer), Bloomberg, Sensor Tower, Wired, and Reuters. All statistics and quotes are drawn from public reporting of Meta’s Q3 2026 earnings call, Walmart’s Masters of Scale Summit announcements, and Recursive’s Flint launch.