On September 21, a pop-up began appearing on the screens of users trying to shop on Amazon with Meta Muse: "Unauthorized AI Agent continuing to access will violate Amazon's terms of use."

This came just 13 days after Muse was officially launched.

Meta's personal AI assistant Muse recently exploded in popularity｜Image source: Meta

Meta launched Muse on September 8, defining it as the "world's first personal AI Agent designed for everyone." Unlike chatbots like ChatGPT and Claude, Muse does more than just answer questions; it can open browsers, log into your email, fill out forms, book flights, and even directly make purchases for you.

The results were immediate, as Sensor Tower data shows that Muse was downloaded over 2.5 million times within 13 days of its launch, topping the U.S. App Store free chart on September 18, surpassing ChatGPT, Gemini, and Claude.

However, faced with the boom of personal AI assistants, traditional internet giants are clearly unwilling to indulge.

01 "Ad-free" Agent

In its rejection letter to Muse, Amazon's statement was polite in wording but firm in logic. Meta never informed in advance that Muse would access Amazon's store; the Agent did not identify itself as an AI while browsing; and it seemed to capture and store user login credentials.

Amazon likened Muse to food delivery platforms and online travel agencies: "These third-party applications, when replacing users to purchase products from other merchants, should operate transparently and respect the services providers' decisions on whether to participate."

The words seemed fair, but the real pain points lay buried in the financial reports.

Amazon's ad revenue last year exceeded $68 billion. This revenue source is very specific; users input keywords in the search bar, scroll through sponsored products, and are guided by recommendation algorithms, eventually clicking buy. The premise of the entire chain is that a person is browsing.

The AI Agent does not browse; it does not scroll pages, look at sponsored products, or get influenced by recommendation algorithms. It only does one thing: find the item the user wants and then check out.

In transactions initiated by the Agent, advertising loses its purpose.

This is not the first time Amazon has taken action against AI shopping Agents. In November 2025, Amazon sued Perplexity, arguing that its Comet browser's AI Agent impersonated ordinary users logging into Amazon accounts to shop. In March 2026, a federal judge approved a preliminary injunction, blocking Comet's access to Amazon's password-protected areas.

AI does not "shop"｜Image source: techspot

However, the story took a turn in August.

The Ninth Circuit Court of Appeals lifted this injunction for interesting reasons: based on existing evidence, it is the users accessing Amazon's system, not Perplexity. The judge's logic was that the Computer Fraud and Abuse Act is essentially an anti-hacker law, and the scenario where users authorize Agents to operate on their behalf falls outside its scope.

After legal tools became ineffective, Amazon turned to a more ancient and effective means—terms of service.

Since July, Amazon has quietly begun a series of "reinforcement" actions, removing specific product names from confirmation emails, making it harder for AI Agents to interpret purchase records; expanding the robots.txt file, blocking 47 AI crawlers. By the time Muse actually arrived, Amazon only needed to present a pop-up to close the door.

On September 10, the court denied Amazon's request for a rehearing in the Perplexity case. However, the judgment left a gap, allowing for litigation claims based on contracts and service terms to proceed.

In other words, Amazon's battle in court is not truly over; it has just changed tactics.

02 Agent's Native Architecture Compensation

On the same day that Amazon pulled the plug, Shopify CEO Tobias Lütke announced an entirely opposite piece of news: Shopify would collaborate with Muse to enable the Agent to complete checkouts directly in Shopify merchants’ stores.

In fact, Shopify's actions began even earlier.

On September 8, the day Muse launched, Shopify added Meta to the "Agentic Storefronts" AI channel. This means all eligible Shopify merchants’ product data is open to Muse by default, with no extra action required from merchants.

The underlying technology supporting this is called the Universal Commerce Protocol, or UCP. This is an open standard jointly released by Shopify and Google in January 2026, defining how AI Agents can discover products, negotiate transaction terms, and complete payments. More than 20 retailers and payment networks, including Etsy, Target, Walmart, Wayfair, Visa, Mastercard, and Stripe endorsed it at launch.

Shopify actively supports the integration of the Muse application｜Image source: Naughton&Bird

The design philosophy of UCP stands in stark contrast to Amazon's defensive posture. It treats Agents as a new distribution channel rather than intruders. Merchants declare which capabilities they support within the framework of the agreement, and the Agent chooses the appropriate payment method for transactions, all while the user remains in the conversation interface. Merchants remain the main party in the transaction, retaining customer relationships, pricing power, and all transaction data.

The "zero-click shopping" that Amazon fears is defined as a new shopping experience within Shopify's framework.

Where's the difference? Amazon's business model is "advertising-driven traffic monetization." Every search, every scroll, every pause by the user is the basis for advertising revenue. The Agent skips all of this, whereas Shopify’s business model is "charging commissions for providing infrastructure to merchants." Shopify does not care where users come from or how they find products; it only cares whether a transaction occurs and whether it uses its checkout system.

So, Amazon closes the door, and Shopify opens it. It's not because one is more enlightened than the other, but because their business models dictate their attitudes towards the Agent.

03 The "Same Wall" Faced by Personal AI

Almost in the same week that Muse was rejected by Amazon, a similar story unfolded in China.

On September 16, the second-generation Doubao phone (Nubia NaviX Ultra) was officially launched, starting at 5999 yuan, with an initial stock of 200,000 units and nearly 400,000 appointments on JD.com.

If you recall the excitement last December, the first-generation Doubao phone came with an AI phone assistant capable of simulating clicks via a GUI, replacing users in operating almost all apps on the phone. The 3499 yuan engineering machine sold out 30,000 units on the first day, with the second-hand market once inflating prices to nearly 8000 yuan.

Then less than 48 hours later, super apps like WeChat, Alipay, Meituan, and Taobao collectively shut down access. Risk control mechanisms precisely intercepted the AI's simulated click actions, with some banking apps even issuing direct pop-up warnings. The Doubao team was forced to take down the operational permissions for financial payment scenarios, turning "fully automatic" into "semi-disabled."

The logic of super apps is identical to that of Amazon; if users become accustomed to having AI complete operations directly, the app would lose user time, advertising exposure, and data sovereignty, degenerating into an indiscriminate backend service provider.

Nubia's "Doubao Second Generation Phone"｜Image source: Nubia official website

The Doubao Second Generation, returning after over nine months, has learned its lesson. The technical route changed from "breaking in" to "knocking at the door," shifting from GUI simulated clicks to MCP and A2A protocol driven approaches. The Doubao team even issued a SAEP automated screen operation declaration agreement, providing apps a 30-day notice period to decide whether to accept AI operations.

But what was the result? Real-machine tests show that mainstream applications like WeChat, Meituan, and Taobao still cannot achieve automation. Currently, the only products that can be invoked smoothly are basically Byte's own products. More notably, Tencent has released A2A assistant capabilities for WeChat in collaboration with Huawei, Honor, Xiaomi, OPPO, and vivo, but this collaboration list does not include ByteDance.

What Doubao and Muse face is actually the same structural problem; AI Agents want to become an intermediary between users and services, but the platforms controlling "user reach" have no reason to give up that position.

The difference on both sides lies in the form of expression. In the United States, Amazon utilizes terms of service and court litigations, with the logic being "you have no right to act on my users' behalf on my platform." In China, super apps use risk control interceptions and ecological alliances, with the logic being "you are not my friend, so you cannot knock on my door."

However, the underlying conflict is entirely consistent: Whoever controls users' attention and purchasing decisions controls the distribution of commercial value.

The emergence of AI Agents is threatening this distribution power.

04 Bans Are Not the Endgame

If looked at purely from the present, the situation for AI Agents is indeed not optimistic. But when the perspective is broadened, a ban is not necessarily the end.

The UCP alliance led by Shopify and Google has already covered more than 20 large retailers and payment networks. Visa, Mastercard, and Stripe are all developing their own Agent payment protocols. If more merchants realize that Agents can bring incremental transactions rather than robbing existing traffic, opening up will be a more rational choice.

Amazon cannot permanently remain the "rebel" either. It has its own AI shopping tools, Rufus, and Alexa for Shopping, but these tools are meticulously designed in an "advertising-compatible" manner, displaying sponsored content while recommending products. Amazon's ban is not against AI shopping itself but against AI shopping that it cannot control.

Consumers also have a crucial variable on their side.

Oppenheimer's survey of 1,500 American consumers shows that only 8% trust Meta to manage their passwords, while 58% are unwilling to share their passwords with any AI Agent.

The 2.5 million downloads of Muse are certainly impressive, but whether it can cross the trust threshold is key to determining how far this path can go.

To some extent, today’s conflict between Muse and Amazon resembles the scene when mobile payments first emerged. Banks and traditional financial institutions once worked hard to prevent third-party payments from entering their territory; the final result was that both sides found a way to coexist, although the balance of power shifted permanently.

The game between AI Agents and super apps might not end in the total victory of either side. A more likely outcome is that Agents learn to "knock on the door," and platforms learn to "open the door," while the rules behind that door will take a long time to negotiate.

However, before such negotiations are reached, every AI trying to "do things" for users must first be blocked outside that door.