Recently, the Doubao mobile phone user base has exploded. Some users of the first-generation Doubao phone (Nubian M153) found that after September 14, they could no longer operate the phone through the GUI, including hailing a taxi, ordering takeout, and cross-end operations. Many users said: “I am used to using AI to control the Doubao phone to hail a taxi and order takeout. I am no longer used to doing it manually.”

The M153, which will be launched in December 2025, sparked discussion as soon as it was launched, and tens of thousands of units were sold as soon as it was launched. Once many users get used to using Agent to operate the mobile phone interface, it is difficult to roll back to the previous state. Some people call it “Agent Syndrome”.

But this time, it was not the super application or the cooperating mobile phone manufacturer ZTE that pressed the pause button on the Doubao phone, but the team itself. This was due to a new rule released by the Doubao phone.

On September 16, Nubia NaviX Ultra, a mobile phone equipped with the consumer version of Doubao Mobile Assistant, was released, priced from 5,999 yuan. Released together with the mobile phone is the official “Screen Automation Execution Protocol” (SAEP, Screen Automation Execution Protocol).

The agreement mentions that starting from September 14, the official will set up a 30-day publicity period. During the publicity period, the APP will not be operated by GUI agents unless it actively declares acceptance. If the APP is not notified by email and does not declare rejection through the SAEP protocol, they will open the ability of the GUI agent to operate the application.On September 17, Doubao Mobile Assistant added in its “Answer to Netizens’ Questions”: After the publicity period, third-party application developers can still declare their refusal, and we will shut down related operations in a timely manner.

“Chinese Entrepreneur” noticed that Doubao mobile assistant can currently provide Doubao smart services, Doubao shopping, recording minutes, work tasks and other AI experiences. At the same time, users can also call Byte-based applications such as Douyin, Feishu, Doubao Office, etc. on their mobile phones through GUI intelligent agents.

With the help of AI and the maturity of GUI agent technology, it has become inevitable for AI assistants to step out of question and answer dialogues, understand mobile phone interfaces, and control across apps. While this capability brings about an efficiency revolution, it also raises a new proposition: when AI agents can respond to user instructions to help operate third-party apps, how to delineate the boundaries of developers, how to protect the will of users, and can multi-party conflicts be resolved without bans and confrontations?

SAEP has made a pioneering attempt, using rules and mechanisms such as pre-negotiation, hierarchical classification, and public disclosure, in an attempt to provide boundary channels for three-party applications. As many scholars believe: in the final analysis, the right to call intelligent agents belongs to the users themselves.

Ge Jiangqiu, associate professor at Fudan University Law School, said: “Dual authorization” of intelligent agents is not the right answer.“In terms of function, the intelligent agent essentially assumes the role of assisting or acting as an agent. Therefore, no matter what technical means or behavioral patterns the intelligent agent adopts, the final legal effect should be attributed to the user.”

Zhai Wei, executive director of the Competition Law Research Center of East China University of Political Science and Law, also analyzed from the perspective of competition law that users have the right to self-determination over their own data—including the right to portability and the right to be forgotten. Interconnection has never been a new topic, but every time technology generations change, it comes back to people with a new look.

“Laws and regulations such as the Personal Information Protection Act have clarified that user data can be transferred between different applications and permanently destroyed.Cross-application access is legally a reflection of the user’s will. Authorization of third-party applications is not a necessary prerequisite here, nor should it become an access barrier unilaterally set by the platform. “

The second half of AI mobile phones is here

Looking back at the history of the development of mobile phones in China, at every node of technological change, from the “Apple tax” to “channel sharing” to the “battle between advertising push and user privacy boundaries”, the three-party game between mobile phone manufacturers, system manufacturers and super applications constitutes the most vivid footnote of the industry.

Entering the AI ​​era, contradictions have reached a new node. The connection between mobile phone AI systems and traditional applications is becoming increasingly urgent. Mobile phone manufacturers such as Apple, Huawei, and Xiaomi are strengthening their terminal-side AI capabilities and launching AI portals with innovative features.

But at the same time, the industry Agent interconnection agreement has not yet reached a consensus. The A2A protocol developed by Google, the MCP protocol promoted by Anthropic, and various Skill plug-ins, as well as the Agent payment and transaction network that people are actively discussing, still lack mutual trust and collaboration standards. This also makes traditional APP application software and Agents open and interconnected, and frequently hit the wall of regulations.

At present, the industry has provided a variety of problem-solving ideas. For example, traditional software opens the backend API and can be called by the Agent. The advantage is stability, but the disadvantage is that it cannot access local data and has a single path. Or, like Meituan, WeChat, etc., independently develop Agent functions, but they can only be limited to this application and cannot be called across domains.

Some AI companies have also tried virtual machines, but they rely too much on cloud resources and have poor stability. Another method is GUI Agent invocation, which simulates user clicks and can access both applications and the local context on the device. The disadvantage is that it is unstable and depends on the improvement of model capabilities.

GUI Agent is not the first of Doubao, but it is the largest player in bringing this capability to the public and completing large-scale iterations. It has also truly advanced mobile phone AI from a “simple chat robot” to a mobile agent that can actually do work. It can be said that it has truly opened the second half of AI mobile phones.

On the first-generation Doubao mobile phone launched in December 2025, users can already experience real agent calling scenarios: comparing prices, ordering takeout, and hailing a taxi with one click. Every step of the AI ​​operation is visualized on the screen. Although it was quickly “blocked” by WeChat, Alipay, game manufacturers and other applications,However, 90% of high-frequency applications on the market have successively supported the GUI Agent operation capabilities of Doubao mobile phones.

“Chinese Entrepreneur” has seen that the functions of the consumer version of Doubao Mobile Assistant, which was officially launched recently, have been improved. Users do not need to take additional screenshots or switch applications, and can let the assistant continue processing tasks based on the current screen content.

For example, in the official video demonstration, the user turns on the camera, points it at the indoor environment, and says, “Based on this decoration style, help me choose a cabinet within 1.2 meters and with a price of no more than 1,000 yuan.” Doubao can combine the screen information and user requirements to find qualified product recommendations on the e-commerce platform.

Task execution is another core capability demonstrated by Doubao Mobile Assistant this time. Taking the travel scenario as an example, if the user says “call me a car to go to the company”, Doubao can confirm the usual address, select the car model and initiate the order.

Robots protocol in the AI ​​era resolves conflicts beforehand

In the past, similar conflicts around the world mostly occurred after the fact. When disputes occurred, they were technically blocked. There was a lack of unified and standardized prior consultation mechanism. In order to completely resolve the concerns of stakeholders, Doubao Mobile Assistant decided to put the brakes on itself first.

The consumer version of Doubao Mobile Assistant first opens its GUI capabilities in Beta form, instead of pursuing full rollout in one step, and prioritizes ensuring stability and usability. The accompanying SAEP agreement is an institutional response to the reality of the industry.

“Graphical interface intelligence is still a rapidly developing cutting-edge capability. Doubao Mobile Assistant hopes to work with partners to build an open and standardized mobile AI application ecosystem.” Beanbao said.

This image uses AI generation technology

The Screen Automation Statement Protocol (SAEP) is regarded by the industry as the Robots protocol in the AI ​​application era. The Robots protocol in the early days of the Internet allowed websites to independently declare whether to allow search engines to obtain content; SAEP returned the choice of “whether to allow AI agents to operate this App” to developers.

SAEP “open-sources” the technical logic, implementation framework, and operation strategies of Doubao GUI Agent, and puts on the table the functional modules of the Agent’s calling applications and which steps are prohibited or allowed to be accessed by developers. It can be said that it considers the interests of developers as the most important consideration. The first page of the agreement mentions that it mainly solves four problems:

First, let the application express operational boundaries:Third-party applications can declare which pages and operations allow Agent access, and which pages or capabilities need to be prohibited from GUI Agent agent operations;

Second, let strategic management make unified decisions:The policy management module integrates the Agent identity, system security baseline, application declaration policy and user authorization status to obtain the final decision for this operation;

Third, make the constraints not easily bypassable:Restrictions are executed by the key capability entrance of the system and not only rely on Agent self-discipline;

Fourth, make behavior traceable:Record who was at what time, to which application interface, what operation was requested, what policy was hit, and the actual results.

More importantly, SAEP has set up a “triple gate” to protect the interests of developers.

比如事前決策. Set a 30-day rule publicity buffer period. During the publicity stage, only operating system applications, byte system applications, and third-party applications with explicit consent will be used. The other applications will not operate by default, leaving a window for learning, evaluation, and feedback for the entire industry.

Be flexible and adjust on the fly.Developers can adjust application authorization permissions at any time, implement refined classification and hierarchical control, and individually prohibit high-risk actions such as deletion, publishing, and sign-in draws, without making an “all or nothing” blanket decision.

Negotiate afterwards to clarify risk stratification and rights and responsibilities.After the publicity period expires, applications that have not expressed their stance will be gradually released according to risk levels; applications that have been explicitly rejected will never perform automated operations.

Looking around the world, major overseas manufacturers are also exploring GUI intelligent agent governance specifications. OpenAI’s Operator and Anthropic’s Claude Computer Use are also based on “screenshot + multi-modal understanding”, and each company has its own rules.

For example, Anthropic stipulates that browsers and trading platforms are only allowed to read, but not modified; the code editor can be clicked, but input is prohibited; other applications are directly given higher operating permissions, and there is no unified negotiation agreement for the entire industry.

In contrast, the SAEP protocol launched by Doubao Mobile Assistant is a valuable attempt to explore further in commercial scenarios in China and even the world.

Does the agent belong to the user or the manufacturer?

After the GUI Agent is implemented, there is still a practical problem that cannot be avoided: how to judge the power of the end user, and whether it can independently determine the permissions of the Agent and super application. In other words, does the GUI require dual authorization (both user consent and platform permission).

Shi Daxiao, associate professor at Fudan University Law School, said:AI agents should be regarded as “digital clones” and should respect the rights and interests of users rather than guarding the walls for the platform.

“The essence of the AI ​​agent is the technical externalization of the user’s will, rather than a third-party subject independent of the user. The humanoid robot is the body, the AI ​​agent is the soul, and the combination of the two is a real avatar.” Shi Daxiao said that this avatar carries the user’s unique resources and preferences, and its cross-end operation is an extended behavior based on the user’s full authorization.

Wu Peicheng, a researcher of the “Hundred Talents Plan” at Guanghua Law School of Zhejiang University, said that “single user authorization” should be adhered to. If “dual authorization” is implemented, it will not only impose many restrictions on users’ daily use, but also significantly increase the development cost of the intelligent industry.

Huang Yinxu, associate professor at the Law School of Renmin University of China and deputy director of the Legal Technology and Social Governance Laboratory, said that as an agent tool for users, as long as it accurately executes the user’s instructions, the corresponding consequences are subordinate to the user in principle. Authorization completed by the user through system-level settings is legally valid, and there is no need to obtain repeated authorization from a third-party platform.

This is destined to be a long-term game. AI mobile phones are not a one-man show that can belong to a single manufacturer, but an ecosystem involving mobile phone manufacturers, large model manufacturers, a large number of third-party App developers, and end users. If we only pursue the rapid development of technical capabilities and ignore the demands of developers, it will be difficult for the ecosystem to go far; if we are overly conservative, it will inhibit AI innovation.

Steve Jobs once gave a popular definition of innovation: Creativity is just connecting things together.The same is true for the logic of AI. Mature components, capabilities, and resources must be recombined to generate new value. Therefore, interconnection is particularly important.