The second-generation Doubao smartphone goes on sale, and it is time for agent-based smartphones to hand in their exam papers.

On September 16, the nubia NaviX Ultra officially went on sale. It is the second-generation "Doubao Phone" jointly developed by ZTE Corporation and ByteDance, officially positioned as an "AI agent phone".

Last December, its predecessor was only an engineering model priced at 3499 yuan with a limited run of 30,000 units, which required snapping up as soon as it was released. Its price was once driven far above the launch price in the second-hand market. The new generation of Doubao Phone starts at 5999 yuan, and the starting price drops to 5499 yuan after national subsidies. Before the new phone was released, the total number of reservations on JD alone exceeded 370,000 units.

There was a nine-month gap between the limited-edition engineering model and the mass-produced commercial product. This time, the question the Doubao Phone needs to answer is no longer "whether the technology can be realized", but whether this new product category can survive in the market.

From Engineering Prototype to Commercial Product: Making Up for Three Core Shortcomings

Judging from the official configuration sheet, the hardware specifications of the Doubao Phone are not particularly outstanding among this year's flagship phones: it is equipped with the 5th generation Snapdragon 8 Elite platform, a 7100mAh battery, a 6.78-inch BOE display, a triple camera system consisting of a 200-megapixel main camera, a 50-megapixel ultra-wide-angle lens and a 64-megapixel periscope telephoto lens. For storage, it carries CXMT's LPDDR5X with a maximum rate of 10667Mbps. It is worth mentioning that this is the first time a domestically produced high-speed memory chip has been applied to a mass-produced smartphone.

What distinguishes this phone from ordinary flagship devices is the independent AI button with fingerprint recognition on the right side of the body. Pressing this button does not launch a certain application, but activates an agent embedded in the bottom layer of the system. It is connected to the Seed full-duplex speech large model, supports continuous conversations that can be interrupted at any time, and can identify more than ten dialects including Southern Min, Hakka, Shanghainese and Cantonese.

Manufacturers summarize this set of capabilities into four phrases: it understands instructions, executes tasks, remembers context, and ensures security.

In actual scenarios, it works like this: say "Help me book a high-speed rail ticket to Shanghai next Friday, and then book a hotel close to Hongqiao Station", and it will automatically complete the entire process of querying, comparing prices and placing an order; say "Summarize what was discussed in this Feishu group today", and it will read the messages, generate a summary and forward it to the designated person; point at a product and say "Help me find the cheapest option", and it will identify the screen content, compare prices and give purchasing suggestions.

Compared with last year's engineering model, the biggest change of this generation is not that it is "smarter", but that it is "more like a mature commercial product", and the changes are mainly reflected in three aspects.

The first aspect is positioning. The previous generation nubia M153 was a technical preview version, targeted at industry practitioners and early adopters, and sold in limited quantities. This generation is positioned as a consumer version, with a full lineup starting at 512GB, four color options, and follows a mass production retail route.

The second aspect is rule setting. The previous generation directly simulated operations by "reading the screen", but encountered risk control from application developers the day after its launch, leading to serious malfunctions.

This generation has completed a full set of rules, launched the Screen Automated Execution Protocol (SAEP), and initiated a 30-day rule public notice period, during which third-party applications can independently decide whether to allow AI to perform screen operations. During the public notice period, in principle, the Doubao assistant only operates system-built applications, ByteDance's own applications, and third-party applications that have explicitly agreed to access. When sensitive actions such as payment are involved, the identity is verified through the fingerprint button, and the agent permissions are graded by layers.

The third aspect is commercialization. The first-generation Doubao Phone was priced at 3499 yuan with a limited run of only 30,000 units, while the new generation has a mass-produced starting price of 5499 yuan, and the stocking scale has been increased to around 200,000 units.

Overall, the essence of this generation is to reorganize the product that "worked technically but encountered obstacles in the ecosystem" in the previous generation into a commercial product that can be placed on store shelves.

The Malfunctions of the First Generation Are the Starting Point for All Actions of the Second Generation

The direct reason why the second-generation Doubao Phone focuses on "completing the rules" lies in several malfunctions of the first generation.

The reason why the first-generation Doubao Phone attracted widespread attention overnight was its radical technical route of screen reading: the agent does not need applications to open interfaces, but views the screen, clicks and inputs like a human, thus completing tasks across applications.

The advantage of this route is immediate, as it does not need to wait for the cooperation of the ecosystem; the disadvantage is also obvious: it bypasses the authorization of application developers, and once the scale expands, conflicts are almost inevitable.

The day after the first-generation engineering model was released last December, some users reported that when Doubao operated WeChat on their behalf, their accounts were prompted with "abnormal login environment", and some accounts were forcibly logged out or temporarily frozen. In Alibaba's applications such as Taobao, Xianyu and Damai, human-machine verification, crashes and forced logouts occurred frequently. Apps such as Agricultural Bank of China and China Construction Bank suspended login and payment on the grounds of "risk environment".

There are multiple reasons for these countermeasures: if Doubao operates on behalf of users, users will no longer open these apps themselves, leading to the loss of splash screen ads, in-platform traffic and user data. What's more serious is that there is a risk of privacy leakage. The risk control of platforms is to protect the bottom line that "users must use the apps in person".

The second generation responds with the SAEP protocol and a 30-day public notice, allowing apps to declare whether they are willing to be operated by AI. From then on, "whether the operation is allowed" is no longer unilaterally decided by Doubao, but confirmed by both user authorization and app authorization.

In fact, this restraint is not entirely a voluntary choice of Doubao. On July 15 this year, the Cyberspace Administration of China publicly announced for the first time the mobile-side generative AI service filing information of seven manufacturers including Apple, Huawei, vivo, OPPO, Xiaomi, Samsung and nubia. What the AI in mobile phones can and cannot do is being incorporated into the regulatory framework. The "rule improvement" of the Doubao Phone includes both passive correction after hitting a wall and active adaptation to the tightening of industry regulations.

However, even if the Doubao Phone has made some improvements in user security and privacy, the problems have not disappeared, and there are at least three unresolved issues at present.

First, the ecological boundary still exists. At this stage, the voice operations of the Doubao Phone assistant are prioritized to cover ByteDance's own products, and only a few third-party applications such as Caocao Mobility are currently connected. High-frequency entry points such as WeChat, Taobao and Alipay are still not within the operable scope. If an agent that "can execute tasks" cannot handle the most commonly used tasks, the ceiling of its user experience is very obvious.

Second, task success rate is a shortcoming. The statement at the manufacturer's press conference is that the end-to-end task success rate exceeds 80%, but this figure comes from the scenarios defined by the manufacturer itself. Looking at the entire agent industry beyond a single product, the China Academy of Information and Communications Technology cited the OSWorld benchmark data at the World Artificial Intelligence Conference, pointing out that the general success rate of agents performing long-link tasks with more than 500 steps is only 18%.

Third, it is still unclear whether the "one-click task execution" is a rigid demand. The hype of the first-generation engineering model in the second-hand market dropped rapidly after its release, and how many of the 370,000 reservations of this generation can be converted into actual sales remains to be seen after the first sales period.

Four Major Technical Routes Are Bet On Separately, and the Previous Consensus Has Loosened

If the Doubao Phone is only regarded as a new product, its sales performance will only affect ZTE Corporation's financial report for one quarter. The significance of this phone goes far beyond that: it is the first time that the "AI agent phone" category has received market testing in a mass-produced form. At the same time, several other manufacturers have bet on different directions.

Honor takes the "embodiment" route. The Robot Phone released on August 12 is equipped with a 4-degree-of-freedom titanium alloy gimbal inside the body, whose camera can rotate, track, and even nod and shake its head, with a starting price of 9999 yuan. Honor's answer is to give the phone a "physical body", allowing AI to move from digital intelligence on the screen to the physical world.

StepVerse takes the "native" route. Its STEX Neo is equipped with a self-developed agent native operating system Step AOS, which reconstructs the system from the bottom layer, has a built-in personal agent Amoo and the end-side model StepEdge, and is also equipped with an interactive secondary screen on the back of the body.

STEX Neo is currently the only agent phone that has passed the L3 intelligent grading test. Its first batch of ecological partners includes well-known platforms such as Meituan, Alipay, Didi and Ctrip. However, this phone has not yet announced a mass production schedule, and the official also admitted that the shipment volume has not been determined in advance.

Huawei continues its self-developed closed-loop route. The Kirin chip, Pangu Large Model and HarmonyOS work in synergy. The HarmonyOS 7 released in September upgraded Celie to a system-level agent, which can perceive more than 200 system-level user data, and call more than 2100 system capabilities, over 2000 HarmonyOS agents and more than 500 selected skills from partners. Huawei's confidence in choosing this route comes from the fact that it controls both the chip and the system.

The four routes give four different answers to "what an AI phone should be". The divergence itself is not a bad thing: it shows that there is no standard answer for this matter yet, and no manufacturer has found a fully feasible path. No one can tell which direction is correct at the moment.

These four routes reflect a fork in the evolution of smartphones: the integration of agent capabilities into mobile phones is a confirmed trend; the real suspense lies in whether it will make the concept of "smartphone" upgraded, or split into an independent new product category.

This suspense cannot be answered for the time being, because the industry has not yet figured out "what an AI phone should look like".

In the past decade or so, the form of smartphones has been almost uniform: a touch screen and a screen full of app icons. Now, agents have become a new entry beyond app icons, but manufacturers have taken different bets on what form this new entry will take.

The same is true at the standard level. The national standard "Intelligent Grading for Artificial Intelligence Terminals" issued and implemented by the Ministry of Industry and Information Technology in May this year divides terminal intelligence into L1 to L4. In the first batch of tests in July, 66 products from 17 enterprises only reached the L3 auxiliary level, and have not yet touched the highest L4 collaborative level.

The market is also changing. According to the forecast of Counterpoint Research in June this year, phones supporting generative AI will account for 45% of global shipments in 2026, and the proportion will rise to 52% in 2027. The market scale is getting larger and larger, but the distribution pattern has not been determined yet.

In the final analysis, the answer must be given by users. In the past, what drove users to replace their phones was better experience brought by larger screens, faster chips and better camera performance. When these three dimensions have reached the ceiling, AI agents are expected to become a new reason for users to replace their phones.

Whether this expectation can be realized depends on two things: first, whether the agent can change from a novel trial function to a daily rigid demand, and second, whether it can leverage the ecosystem of super apps. The former determines whether users are willing to replace their phones for it, and the latter determines how large the scope of "tasks it can execute" is.

At present, the Doubao engineering model has proved that "it can be realized technically", and now it is the market's turn to prove that "it is worth buying". The answer sheet of the Doubao Phone has been submitted, and the ones who grade it are every real user in the next few months.

This article is from the WeChat official account "Deep Gaze Finance" (ID: chutou0325), the author is a follower of the AI phone track, and it is published with authorization from 36Kr.