DouBao Mobile increased its inventory from 30,000 to hundreds of thousands of units, but removed the GUI capabilities of the previous generation—no longer reading screens or simulating taps on top apps like WeChat. Instead, it now relies on the MCP protocol to wait for apps to voluntarily grant permissions. Unlike DouBao and Nubia, JieYue XingChen adopted a full-stack approach to launch STEPX Neo and its operating system, while Honor remains focused on its mobile OS and has partnered with Alibaba and others. These three paths all point to one core issue: commercializing Agent phones requires super apps to open APIs, rather than attempting to bypass app boundaries. With the emergence of protocols like MCP, App Intents, and AHA, the industry is shifting from “bypassing apps” to “waiting for apps to open their doors,” with super apps increasingly serving as the backend service systems for Agents, while GUI continues to handle long-tail scenarios.

Author and source: New Position Pro

The closer it gets to commercialization, the more conservative the capabilities become.

According to reports, the new generation of DouBao phones has increased their inventory from 30,000 units to hundreds of thousands, but the GUI capabilities that drew the most attention in the previous generation have been removed. The new product will no longer read screens or simulate taps on top apps such as WeChat, Taobao, and Alipay; DouBao phones can only integrate with apps that proactively offer MCP services and grant corresponding data and operation permissions.

Before and after the adjustment of the route in Doudou, the Agent phones converged toward the front of the stage.

On July 13, Jiepoin AI released STEPX Neo, the Agent Operating System Step AOS, and the personal agent Amoo; on July 15, the Cyberspace Administration of China announced the registration of seven mobile-side generative AI services; on July 18, Honor showcased the Robot Phone at WAIC and further advanced MagicOS toward an Agentic OS.

Three companies have chosen three different paths. DouBao partnered with Nubia to integrate model capabilities into an existing smartphone brand; Jieyue developed its own models, operating system, and hardware; while Honor focused on maintaining its smartphone operating system and then brought in external model and service partners like Alibaba.

Three products at different stages are simultaneously pushing the industry toward the same question: Large model companies understand users, and smartphone manufacturers control devices, but the social, payment, shopping, and local life services people use daily belong to a different set of companies. The first step toward commercializing an agent-enabled phone is not to gain more permissions, but to acknowledge that some permissions never belonged to it in the first place.

Typically, as a technology matures, its features become increasingly complete, but Agent phones are undergoing the opposite process.

Everyone, get on your phones first.

In the years following the emergence of large models, AI was integrated into glasses, headphones, pendants, recording cards, and other new devices. These devices dominate certain scenarios but struggle to simultaneously possess what smartphones already offer—stable, large-scale user bases, continuously online sensors, personal context such as photo albums, calendars, contacts, and location data, payment and communication capabilities, and the channels and supply chains needed to deliver products to everyday users.

Large model companies initially chose smartphones because, at this stage, only smartphones can integrate users, permissions, payments, and services onto a single device. At the July 13 launch event, Yin Qi stated directly: Any model company that doesn’t own an application sufficient to sustain it will struggle to survive in the long term. This also explains why Jiepai is simultaneously launching Step AOS, Amoo, and STEPX Neo alongside its foundational model.

Model APIs can generate usage revenue but do not directly establish user relationships. Standalone AI applications can acquire users but struggle to obtain full device permissions and personal context. For an agent to complete a real task on behalf of a user, it needs to know the user’s identity, location, schedule, and preferences; it must invoke third-party services; and it must receive feedback from the task outcome. Terminals connect all these components together.

Step Function’s expectations for STEPX Neo are not to set sales targets first and then sell. Yin Qi stated that the current product is approximately 70% complete, with the remaining aspects—such as licensing, trust, cloud-device division of labor, and specific use cases—to be refined through user feedback.

This is how the model company enters the mobile market: first, build its own device, then use it to validate the relationships between the operating system, the model, and the services.

DouBao’s approach is lighter: ByteDance provides the model and agent capabilities, while ZTE and Nubia handle hardware, manufacturing, sales, and after-sales service. Compared to building a supply chain from scratch, this path is faster; compared to merely selling models to smartphone manufacturers, it allows deeper involvement in the product experience.

The goals of both parties are not entirely aligned: ByteDance needs real users, task data, and new service entry points, while Nubia needs to secure a perceptible differentiator in a highly homogeneous Android market. Model companies aim to build long-term services, whereas phone companies must first sell devices. The former focuses on usage, memory, and subscriptions; the latter must also account for inventory, distribution channels, and hardware gross margins.

Compared to software companies, smartphone manufacturers’ reasons for entering the Agent space are more akin to a changing cost sheet. Starting at the end of March, prices for storage and other core components began to rise, prompting most Android manufacturers to increase prices, further dampening users’ willingness to upgrade. Preliminary data released by IDC on July 14 showed that smartphone shipments in China during the second quarter of 2026 totaled approximately 66 million units, a 4.3% year-over-year decline—the fifth consecutive quarter of decline; shipments for the first half of the year reached approximately 134 million units, down 4.2% year-over-year.

Over the past few years, smartphone manufacturers have sequentially justified premium pricing with advancements in imaging, fast charging, foldable screens, and satellite communication. This round, it’s Agent’s turn. A phone running an Agent continuously requires larger memory to store models and context, a more powerful chip to perform on-device inference, and the ability to orchestrate complex tasks between the device and the cloud. These investments can serve as justification for price increases and will continue to raise hardware and service costs. However, if users ultimately gain only an additional interface for summaries, photo editing, and Q&A, it may be insufficient to drive a new upgrade cycle.

The agent has therefore been given a dual mission: to help smartphone manufacturers establish new product premiums, and to prove that the experience it delivers justifies the additional costs.

From bypassing the app to waiting for the app to open

Mobile devices have another layer of restriction: existing operating systems are built around the concept of "humans using apps." Each app has its own data, accounts, and permissions, and the system uses sandboxing to prevent them from crossing boundaries. Payment, social, and banking apps each have their own authentication and risk control mechanisms.

The agent works in the opposite way: when a user says, “Help me arrange a business trip,” it must check the calendar, confirm traveling companions, search for flights and hotels, and then initiate a payment. Although there is only one task, the required data and services are scattered across multiple apps. While the model can break down this task, it cannot determine by system permissions alone how apps like Ctrip, Fliggy, and the bank should coordinate.

The smartphone contains everything an agent needs and consolidates all the boundaries formed over more than a decade of mobile internet, making it the most practical entry point—and also the most difficult to transform.

The GUI demonstrated that DouBao can quickly prove AI is now capable of using apps just like a human. This technology does not require apps to provide specialized interfaces; instead, the model observes screens via screenshots, identifies buttons, input fields, and pop-ups, then simulates taps, swipes, and typing—any task a user can perform, the AI can theoretically attempt as well.

There are millions of apps on mobile devices, and waiting for developers to update each API individually would take a very long time; having AI directly understand existing pages allows for rapid coverage of a wide range of scenarios. Users can also clearly see AI in action—opening one app, switching to another, and sending the results back.

However, for app developers, user authorization for AI to read the screen does not equate to platform approval for the AI to bypass its own APIs to perform actions. Apps cannot determine why the AI is reading the page, what specific information it needs, or distinguish between user-authorized assistants and bulk automation tools.

Technically, Doubao could click on behalf of users; commercially, apps have not accepted external agents operating freely within them. WeChat, Taobao, Alipay, and several banking apps have successively restricted these capabilities. Public information shows that Doubao subsequently disabled its ability to operate WeChat and banking and internet payment apps.

MCP brings this relationship back to the table. Under the new scheme, applications must actively declare which data and actions they are willing to expose. They can allow agents to query orders but not modify them; permit entry of purchase information while leaving payment confirmation to the user; or enable access to a specific service without granting full account data to the system.

The scope of AI-initiated actions has been changed from being autonomously determined by the AI to being jointly agreed upon by the model provider, smartphone manufacturer, and application. Agents now have more stable execution paths, while applications regain control over permissions, risk management, and data boundaries.

From the outset, STEPX Neo's path has been more aligned with this approach. Its initial partners include Alipay, Baidu, Meituan, JD.com, Didi, Ctrip, and Amap, with application capabilities integrated via protocol interfaces rather than GUIs.

During the WAIC demo, the user told Amoo: “Please deliver the contract to the company—Frank needs it urgently.” The system retrieved the address from contacts and context, ordered a delivery via Meituan, and generated an order pre-filled with the item, cost, and delivery time. Most of the process was completed within Amoo’s interface; when it reached the payment stage, the task paused on the Meituan page, awaiting user confirmation. The system agent understood the intent and completed information organization, while Meituan retained responsibility for fulfillment and payment, and the user retained final approval authority.

Apple is also incorporating these boundaries into its system. According to Apple’s 2026 updated developer documentation, apps must actively declare via App Intents the actions and data that can be invoked by Apple Intelligence and Siri; private content within the app interface will not be automatically exposed to the system, and sensitive or disruptive actions may require additional user confirmation.

Alipay refers to this cross-device collaboration interface as AHA. On July 15, "Abao" connected with OPPO's "Xiao Bu Assistant," with Xiao Bu handling request understanding and Abao invoking nearly 200 in-house Alipay lifestyle services, while key authorizations and payments still require user confirmation.

MCP, App Intents, AHA, and other structured interfaces are becoming the solutions offered by different companies. They are not the same protocol, but they address the same problem: enabling apps to proactively determine which capabilities can be invoked by agents, rather than having system agents enter apps, read pages, and perform actions on their own.

Within the system, alarms, albums, calendars, and files can be broken down by smartphone manufacturers into atomic capabilities; payments, social, e-commerce, and local life services are exposed by apps through official interfaces; unadapted long-tail apps may still rely on GUI supplements.

Therefore, the GUI will not disappear immediately. It still serves as a shortcut for legacy applications and low-risk tasks. However, the closer we get to payments, social interactions, and transactions, the more the Agent needs a path that is recognized by applications and can be audited and held accountable.

The prototype focuses on whether AI has a hand; for mass-produced products, the priority is determining how far that hand can reach.

The app is willing to fulfill but unwilling to be assigned.

The super app certainly does not resist agents.

Alipay is developing Abao, Alibaba is integrating Qwen into Taobao, WeChat is beginning limited testing of "Weixiao," and smartphone manufacturers are upgrading their voice assistants into system agents. Each company is vying to be the first to respond when a user voices a request—and to determine which service to invoke.

On May 11, Alibaba integrated the Qwen app with over 4 billion products on Taobao and Tmall. Users can now perform product searches, comparisons, order management, logistics tracking, and after-sales services within Qwen; the Qwen shopping assistant has also been launched within the Taobao app. Alibaba is willing to let an agent make purchases on behalf of users, but this agent first belongs to Alibaba, the products come from Taobao and Tmall, and payments and after-sales services remain within Alibaba’s ecosystem.

The past mobile internet business chain began when users opened an app. Platforms influenced choices through home pages, search, recommendations, and ads, then completed transactions and subsequent marketing. System agents aim to move the decision-making earlier: users no longer need to open Meituan, Taobao, or Ctrip—instead, they simply say, “Order me a coffee” or “Book a hotel nearby.” The agent understands the request, compares services, and returns results, with the app stepping into the background to complete the order.

For users, it’s just a few fewer steps. But for the platform, it may mean losing homepage exposure, search ads, recommendation rankings, and behavioral data. The app may still complete the transaction, but it no longer necessarily knows where the transaction began. A user’s coffee could come from Taobao Flash Sales, Meituan, or the brand’s own mini-program; a hotel room could be booked through Ctrip, Fliggy, Meituan, or direct channels. Which provider the system agent ranks first determines where the order, commission, and user data flow.

Therefore, the super app will not unconditionally hand over this control to the mobile operating system.

The more likely future structure will involve a division of labor between system agents and app agents: system agents, closer to the device, understand the user’s location, schedule, contacts, and current intent; app agents, closer to the business, manage products, accounts, social relationships, payments, and fulfillment. Model providers handle understanding and planning, smartphone manufacturers manage system permissions and end-cloud orchestration, and users confirm actions at key points such as payment, deletion, and message sending.

Honor summarizes it as "one master, multiple specialists." In its Agentic OS framework announced on July 18, a portable master agent understands the user and collaborates with vertical agents, end devices, and cloud-based models. Honor also announced a collaboration with Alibaba’s Qwen to develop model solutions tailored for mobile scenarios.

This structure appears more acceptable to all parties, as no single company needs to surrender its entire capabilities. Smartphone manufacturers retain access to the system interface; model companies enter user workflows; internet platforms continue to control services and transactions; and users gain a unified interaction interface.

However, the interface is merely a technical protocol, not a commercial agreement. There are no existing answers regarding which service the system agent should prioritize recommending, whether the app pays for system distribution, whether smartphone manufacturers can participate in revenue sharing, or who bears the cost of model invocation.

Long-term memory adds another layer of competition: whether contacts, consumption preferences, and historical tasks are stored on the phone, the model’s account, or within the app itself will determine which relationships users can take with them after switching devices or services. The interface determines what the Agent can access; memory determines who users find harder to leave.

The mobile internet did not eliminate the previous generation of infrastructure, but it changed how users access services. Similarly, the shift brought by agents may not eliminate apps, but rather move the first step of user-initiated tasks toward the system level.

For a super app, what truly needs to be protected is ranking, trading, and user relationships.

In conclusion

Agent-enabled phones won't quickly converge on a single path. In the short term, three approaches will coexist: AI model companies partnering with phone brands, AI model companies building end-to-end terminals, and traditional phone manufacturers maintaining control over the operating system. Ultimately, all paths will need to negotiate with the same set of super apps.

The app's icon and interface may be opened less frequently by users, but the platform's control over products, social relationships, payments, transactions, and risk management will not disappear. Instead, these functions are more likely to shift from front-end mobile apps to backend agent systems; GUIs will not vanish immediately. They will continue supporting legacy applications with incomplete interface upgrades, while high-risk tasks such as payments, social interactions, and transactions gradually transition to protocol interfaces, tiered permissions, and user confirmation mechanisms.

The New Position holds that the key metrics to watch are how many cross-app tasks can be completed without user intervention, which capabilities WeChat, Taobao, Meituan, and Alipay have opened up, how much cloud computing power an active user consumes, and whether they are willing to pay for these services.

The "hand" that DouBao reclaims is where this business truly begins.