Editor’s Note — DCC.
What the author is reading. Part I, item (4) of MIIT’s Notice on Strengthening the Network Access Management of Agent Terminal Equipment (关于加强智能体终端设备进网管理的通知, 工信厅信管〔2026〕42号), published November 6, 2026. DCC’s full translation of the Notice is on the law page, MIIT Agent Terminal Notice, and DCC’s own map of its obligations is in MIIT Puts AI-Agent Phones Under Network Access Management.
The term. “Dual authorization” (双重授权) comes from Chinese unfair-competition case law on platform data: a third party that takes user data through a platform needs the user’s authorization and the platform’s. The author applies it to an AI agent operating someone else’s app.
How to read it. This is a first-person opinion piece, written on the day the Notice came out. The author is openly disappointed by the rule. The translation is complete and keeps the author’s voice. Product facts and the account of the US litigation are the author’s; DCC has not verified them independently.
On November 6, 2026, MIIT published the Notice on Strengthening the Network Access Management of Agent Terminal Equipment (the “Notice”), which states: “Without a declared consent, an agent shall not forcibly invoke or call up the app.”
This means that before a phone agent operates an app on a user’s behalf, it needs the app’s consent as well as the user’s authorization.
With that, China’s phone agents have formally entered the era of dual authorization (双重授权). I find myself a little let down.
I. Who is affected most?
The “agent terminal” in the Notice is a telecommunications terminal, such as a phone, tablet or wearable device, on which an agent is pre-installed.
Products like the Doubao phone fall squarely within this. They therefore face the requirements on agent adaptation verification, network access testing, reporting of major changes and OTA upgrade management.
But the rules on the distribution of agent apps and on invocation are not limited to the “pre-installed” scenario.
(4) Improve the distribution management mechanism for agents. App distribution platforms shall promote the proactive declaration by app developers, by electronic signature, of whether they consent to invocation by agents; without a declared consent, an agent shall not forcibly invoke or call up the app. App distribution platforms shall publicly display the agent adaptation verification list to users on the distribution page, and shall not provide users with download and installation services for agents that are incompatible with their device model. App distribution platforms shall strengthen patrol inspection of the adaptation verification status of listed agents, and shall take measures such as rectification within a time limit and removal from the platform against agents that have not completed adaptation verification.
On the wording of the provision, a phone agent app that is distributed independently may also be caught by these requirements. One cannot say that item (4) has nothing to do with it merely because it is not pre-installed software.
That said, most phone agents today are more like a “remote control”. The user gives an instruction on the phone, and the task is actually handed to a virtual machine in the cloud or to the user’s PC. Because of operating-system restrictions, they can hardly read WeChat, Taobao and other apps on the phone directly and then simulate the user’s taps. In theory, then, the Notice has little effect on agent apps of this kind. There are technical exceptions, of course: once the user turns on accessibility services and similar permissions, an Android app may be able to read the interface and carry out a certain amount of cross-app operation.
So the form most affected by this Notice is still the one represented by the Doubao phone: an agent pre-installed on the terminal that can use system permissions to operate other apps directly. For WorkBuddy-type apps, the cross-app operation actually runs on the PC, and the effect on them is limited too.
The more interesting case is the cooperation between Apple and Qwen (千问). In the first batch of Apple Intelligence features that Apple has announced for China, Qwen is connected to Siri, Writing Tools and Visual Intelligence in specific scenarios, to help answer questions, understand text and images, or generate content. This update does not include Apple’s next-generation Siri AI features.
The China features made public so far appear to be specific capabilities inside the system: question answering, rewriting, image understanding and photo processing. Qwen cannot yet read the interface across apps on its own or operate other apps in sequence. So Qwen’s entry into Apple Intelligence cannot simply be equated with the Doubao phone’s capability, a system-level GUI agent invoking apps.
If Qwen is later integrated further into Siri AI, and can invoke other apps on its own and complete cross-app tasks on the user’s instruction, it will very probably be an agent within the meaning of the Notice, and the invocation may fall under the adaptation and declaration mechanism in item (4). It will then be very interesting to see whether the major apps dare to refuse access by Apple’s Siri AI.
II. Why I support Doubao’s experiment
When Doubao launched its SAEP protocol, it gave third-party apps a 30-day public notice period. After the period ended, apps that had not expressly refused would be opened up step by step, by risk tier. This was an attempt at opt-out: developers may choose to leave, and if you do not leave you are taken to have agreed. Many people questioned it: is this not implied consent?
In my heart, though, I supported their experiment. See my earlier piece, When an AI Phone Operates Apps for You, Who Decides? (AI手机替你操作App,谁说了算?).
At least before the Notice was issued, I had not seen a general rule requiring any agent that operates any app on a user’s behalf to obtain the target app’s express consent first.
I prefer to understand Doubao’s attempt as a boundary-declaration mechanism similar to the Robots protocol: publish the rules first, let developers express refusal and draw limits, and then explore the service within those boundaries.
It is also easy to understand commercially. Negotiating cooperation one company at a time means talking business terms, dividing liability and signing contracts, and after all that time the deal may still not be done. Put forward a set of rules first, keep a channel for refusal open, try to get the product working, and act first.
My support has a selfish side as well. I do not want to buy an AI phone one day and find that this app cannot be touched and that function cannot be used.
But reality is hard. On the refusal list published when the notice period ended in October, there were already a great many commonly used apps: WeChat, Taobao, Alipay, Meituan, Xiaohongshu, Kuaishou and others.
The traffic logic behind it is not hard to understand either. If users hand their needs to an agent and no longer browse recommendations, pick products and compare services themselves, the attention and ad impressions that used to stay inside the app may shrink. Going further, the agent may also come to hold the power of “whom to recommend to the user”, and become a new traffic entry point. That touches the platforms’ existing commercial interests.
What really worries me is what follows. If the most commonly used apps all refuse to open up, each for its own interests, how many cross-app tasks can an AI phone still complete for its user?
III. History is strikingly similar
This reminds me of the dual authorization principle in Weibo v. Maimai. In case (2016)京73民终588号, the Beijing Intellectual Property Court stressed that a third party obtaining and using user information through an open platform cannot bypass user authorization and platform authorization. Maimai’s conduct was found to constitute unfair competition; the court upheld the first-instance judgment, ordering it to cease the conduct and to pay RMB 2 million in economic loss plus reasonable expenses. Beyond user authorization, platform authorization was treated as a separate requirement.
Of course, “dual authorization” did not become the only line of adjudication in the field of data circulation. The Supreme People’s Court later released Guiding Case No. 263 on data rights and interests, (2019)沪73民终263号, which makes clear:
Where a network platform provides users with a linked-account service and, upon the user’s authorization, transfers data the user has obtained on the linked network platform, thereby making it convenient for the user to process that data within a reasonable scope, and does not disrupt the order of market competition, the conduct does not constitute unfair competition.
— Supreme People’s Court, first thematic batch of guiding cases on the judicial protection of data rights and interests
If platform consent becomes a general precondition for data to flow, the platform may hold a de facto veto. Between that and the aims of promoting the circulation of data as a factor of production and realizing the value of data, there is a conflict that deserves to be taken seriously.
Now that we have arrived at agents, the familiar question appears again. A user wants to use a new tool to do something the user could have done anyway. Should an app be able to decide, with nothing more than “I do not consent”, that it cannot happen?
IV. What if the agent is the user’s tool?
The US case Amazon v. Perplexity gave me a very suggestive angle.
The dispute concerns the assistant in Perplexity’s Comet browser. On the technical architecture the court examined, the user gives an instruction on the user’s own computer; the assistant reads the page information in the browser; Perplexity’s servers take part in the analysis and return operating instructions; and in the end it is the browser on the user’s computer that accesses Amazon. One of the points in dispute was whether that access should be attributed to the user, or to Perplexity as the provider of the assistant.
On August 4, 2026, the Court of Appeals for the Ninth Circuit vacated the earlier preliminary injunction. On the claims under the Computer Fraud and Abuse Act and others, the court held that the one accessing Amazon’s computers was the user, completing the operation with the help of the assistant, and that Perplexity could not be found to have itself carried out the access the statute refers to merely because it received screenshots and supplied instructions. The ruling was confined to the statutory question of “access” on the facts at that time, and the court expressly left room for Amazon to manage access through its terms of service with users.
I like the question it put very much: who exactly is using this tool?
If the user can already access their own account, view the page and complete an operation, and merely chooses to let an agent do it for them, why can this not be understood as the user using a tool to carry out the user’s own intention?
Coming back to phone agents, I think the same question is worth asking. The technical architecture of the Doubao phone differs from that of Computer Use, and the permission chain differs, but these differences are not enough to rule out the “user’s tool” reading automatically.
That an app can more easily express and enforce a refusal at the system layer shows that it can more easily control whether the operation takes place. Whether it ought to have such broad control still needs a separate justification.
As I see it, drawing the scope of authorization, operation and liability concretely may preserve more of the value of this technology than first letting each app decide whether to allow the agent in. See my earlier piece, AI Phones, Different Roads to the Same End: The Compliance Puzzle of Doubao and Gemini (AI手机的殊途与同归:豆包与Gemini的合规迷思).
V. My prediction
How much this Notice will finally affect the industry is still hard to say. For the leading apps that long ago refused in express terms, the new rule changes nothing. The more direct change may come for the apps that have not yet taken a position: the room there once was to try opening up after a public notice period has been squeezed out by a clearer requirement of declared consent.
I am still a little let down. The AI phone I was hoping for is a tool that can get things done for me across apps. I do not want to need Tencent’s agent to reach WeChat and Alibaba’s agent to reach Taobao.
Here is my irresponsible prediction: if practice shows that general refusal by the platforms makes it hard for a phone-agent ecosystem to form, the rules and the ways of cooperating may yet go on being adjusted.
That Taobao was later able to connect to WeChat Pay lets me keep a little hope. The boundaries between platforms, however solid they look, are not unchangeable forever.
This does not mean agents will necessarily take the same road. But at the least, a door that is closed today will not necessarily stay closed.
DCC’s reading
The notes below are DCC’s, not the author’s.
- The author’s scope argument is textual, and it holds on the text. Items (2) and (3) of the Notice are addressed to terminals with a pre-installed agent and their manufacturers. Item (4) is addressed to app distribution platforms and speaks of “agents” at large. Counsel for a standalone agent app should not assume the invocation rule stops at pre-installed products.
- “Dual authorization” is the commentator’s label. The Notice does not use the phrase, and MIIT’s own interpretation presents the invocation rule as user protection. See MIIT’s four-question interpretation.
- The two Chinese authorities cited are unfair-competition rulings, not agent rulings. Weibo v. Maimai and Guiding Case No. 263 were decided under the Anti-Unfair Competition Law on data taken from, or moved between, platforms. Neither concerned an agent operating an app’s interface. DCC’s translation of Guiding Case No. 263 is in SPC Guiding Cases 262–267.
- For the fuller account of the SAEP refusals, including the count of 155 apps and the structure of the protocol, see Li Wenlong’s essay, The Agent Authorization Standoff After SAEP.
— 数据何规, 中国手机Agent进入“双重授权”时代? (Has China’s Phone Agent Entered the Era of “Dual Authorization”?), November 6, 2026. Full translation by DCC. Original (Chinese).
Product descriptions, the refusal list and the account of Amazon v. Perplexity are as stated by the original author; DCC has not independently verified them.
— Not legal advice.