Author: Lu Yao | Editor: Liu Wei
In September this year, one of the most prominent changes in the smartphone industry remains the foldable form factor.
Huawei continues with its tri-fold design, and Apple has finally unveiled its own foldable smartphone. Manufacturers are still grappling with a familiar question: how to fit a larger screen into a limited device body.
Another shift is happening within the screen itself. Large models are stepping out of the chat box and integrating into the smartphone OS.
In the past, looking up information, booking tickets, and scheduling required users to find the right App and operate step by step; now, AI (Artificial Intelligence) is beginning to take over some of these tasks.
While foldable screens change the physical form of the smartphone, Agents may change the way humans interact with it.
Past voice assistants were more like remote controls; for example, "open the camera" corresponded to a single action. An Agent, however, deals with a sequence of actions, such as "schedule next week's review meeting, organize the materials, and send them to the attendees." The AI needs to judge what to do first and which tools to invoke, and it must also know how to proceed if something goes wrong midway.
The Nubia NaviX Ultra, unveiled on September 14 and equipped with the consumer version of the Doubao Phone Assistant, is attempting exactly this. It places AI closer to the system level, hoping that in the future, when users pick up their phones, they won't have to complete every step manually.
This doesn't sound overly complex. The real challenge is whether the smartphone can actually get the job done.
01Models Are Getting Smarter; the Next Step Is Getting Things Done for Users
Over the past two years, model companies have been seeking destinations beyond the chat box.
OpenAI acquired the io team, co-founded by Jony Ive, for approximately USD 6.5 billion, and began developing its own AI hardware; Google continues to advance Agent capabilities within Android, enabling Agents to invoke specific functions within applications through AppFunctions. Similar attempts are also emerging in China, with the Doubao Phone Assistant choosing to collaborate with smartphone manufacturers, while StepFun has launched its own AI terminal ecosystem.
While their approaches differ, the problems they encounter are quite similar: models are already very good at answering questions, so can they now start doing the actual work?
Take booking a flight, for example.
It is no longer difficult for AI to tell you how to buy a ticket. Actually booking the ticket is a completely different matter. It needs to know the date, departure location, and passenger information, and with the user's authorization, log into the account, select the flight, fill in the necessary details, and finally pop up a prompt for the user to take over and complete the payment.
None of this happens within a chat box, yet almost all of it takes place on the smartphone.
A vast array of personal digital services, such as communication, shopping, maps, and calendars, are concentrated on the smartphone. As AI transitions from giving advice to getting things done, the smartphone naturally becomes the next stop.
Smartphone manufacturers face a different challenge.
Faster chips, brighter screens, and slightly better cameras are certainly valuable. But as smartphones have evolved to their current state, these hardware parameter upgrades make it increasingly difficult to create a noticeable distinction with each new generation.
AI provides a new reason to upgrade. Compared to "equipped with a large model," the value that users can more easily grasp is that they will no longer need to operate step by step for certain tasks in the future.
Model companies want AI to enter real-world tasks, while smartphone companies want to find new points of differentiation. Thus, the two sides hit it off immediately; such deep collaboration can also bring a better experience to users.
The collaboration between Nubia and the Doubao Phone Assistant can be understood in the same way.
Nubia is responsible for the complete device and hardware, while the Doubao Phone Assistant provides model understanding and Agent capabilities. Previously, this solution underwent a technical preview on the Nubia M153 engineering prototype.
However, with the NaviX Ultra, the evaluation criteria have changed. A smartphone actually sold to consumers needs to prove that it is "easy to use and can be used frequently."
02Operating the Smartphone Is Not Hard; the Challenge Is Making Users Feel Safe Letting Go
Imagine a very common scenario.
A meeting has just ended, and you say to your phone, "Organize the meeting recording just now, generate a document, add next Wednesday's review meeting to the calendar, and send it to the attendees."
To actually execute this command involves a rather long chain of processes.
First, the AI must understand.
The NaviX Ultra is equipped with a dedicated AI button and also supports interaction methods such as voice, proximity speaking, and AI earphones. Features like continuous dialogue, natural interruption, and dialect recognition essentially solve the same problem: users can just speak normally without having to learn how to talk to a machine first.
Next, the AI must start working.
It needs to organize the recording, generate a document, create a schedule, find contacts, and send the file. Behind these different steps lie different system capabilities and Apps. The truly critical capability of an Agent is that it must not only know how to operate the screen, but also understand what is currently happening and know what to do next when encountering logins, pop-ups, page changes, or execution failures.
Memory solves another problem.
An assistant that treats you like a stranger every time you meet can hardly be called an "assistant." If the AI already knows who the attendees are and where the user habitually stores meeting materials, there is no need to ask again next time. Users can proactively have the phone remember information displayed on the screen or explicitly told to it, forming a memory for future retrieval.
However, the more it knows, the more critical the permission issue becomes.
Organizing meeting minutes can be done automatically, but sending them out should not necessarily be automated. The same applies to payments, deleting files, sending messages, and reading information.
According to Nubia's solution, high-risk operations require user confirmation and, when necessary, user takeover; the inclusion of fingerprint recognition in the dedicated AI button is also intended to combine wake-up and identity verification.
Agent smartphones are hardest to balance here: they must be proactive enough, but not overly proactive.
Therefore, when evaluating such products, just looking at how many tasks they can perform is not enough. More importantly, one must consider the task success rate, the number of manual takeovers, and exactly how long the user needs to monitor it.
If the AI takes 90 seconds to complete a task while a human would only need 60 seconds manually, the AI hasn't necessarily lost. As long as the user can do other things during those 90 seconds, it still saves human effort.
Conversely, if the AI gets stuck every two steps, requiring the user to constantly confirm, correct errors, or even take over, even if it only takes 40 seconds, we have merely shifted from being "smartphone operators" to "AI supervisors."
What the Agent should truly save is the attention the user invests in completing a task.
This also explains some of the hardware design of the NaviX Ultra.
The 7100mAh battery corresponds to continuous standby, communication enhancement relates to the stability and continuity of interactions with the agent, and the 200-megapixel sensor, besides taking photos, can also serve as an entry point for the AI to acquire visual information; the dedicated AI button provides a more direct wake-up and authentication method.
None of these hardware components are novel in themselves. What is new is that they are being recombined around the goal of "keeping the AI working continuously."
As for whether the battery life is sufficient, whether it is stable under weak network conditions, and how high the success rate is for complex tasks, these ultimately have to be answered through actual usage.
03From Operating Apps Manually to Letting Agents Help Operate Them
Another change brought by Agents is that the way users interact with Apps is evolving.
In the past, most smartphone tasks started with the user operating manually. Once a need arose, they would first find the corresponding App, enter the page to look for the function, fill in the information, and complete the task step by step.
With the emergence of Agents, users can first express what they want to accomplish, and then the AI helps complete the subsequent operations.
For example, "Help me buy a ticket to Shanghai for tomorrow afternoon."
Here, the user does not specify which App to open or which page to enter. The Agent can find the corresponding service based on the task, complete the search and information filling, and hand back the parts that require the user's decision or confirmation.
Flight tickets are still provided by airlines and travel platforms, food delivery still comes from restaurants and delivery platforms, and content is still hosted by various applications. The change is that some of the searching, comparing, filling, and page operations that previously required the user's personal involvement can now be handed over to the Agent.
Google's AppFunctions is attempting to solve a similar problem: developers can open up application capabilities such as calendars, notes, and task management for Agents to invoke.
Therefore, Agent smartphones are redefining the interaction between humans and smartphones.
In the past, users often had to figure out "what steps I need to take" first; now, they can first say "what I want to accomplish," and then the Agent helps invoke the appropriate applications and functions.
As Agents gradually take on more operations, a new dimension will be added to smartphone competition: when a user has a need, who will they hand the task to first?
Because whoever understands the task first has a better chance of helping the user organize the subsequent operation path.
This does not mean that Agents can make decisions entirely on behalf of users.
Agents can perform fewer operations on behalf of users, but the final right to choose should still remain in the hands of the user.
04Smartphones Begin to Complete "Tasks"
On September 16, the Nubia NaviX Ultra officially went on sale. According to official data from Nubia, sales exceeded CNY 100 million within one second of launch.
For the Nubia Navi X Ultra equipped with the consumer version of the Doubao Phone Assistant, the real test has just begun.
The "AI operating the smartphone," which previously remained on engineering prototypes, has now entered officially sold flagship devices and must face network fluctuations, application updates, and various different user usage and scenario needs.
For Nubia, a more critical issue is whether it can gradually get users accustomed to handing over some smartphone tasks to the Doubao Phone Assistant.
Over the past decade or more, smartphones have developed a very mature usage pattern, and this pattern will not disappear simply because of the emergence of Agents.
The change is happening to the people operating them.
Users still use familiar applications and services; it's just that more and more often, they can first tell the phone what they want to do, and then let the AI help complete the subsequent steps.
From users operating smartphones themselves to Agents helping users operate them, the rules of smartphone competition are also changing accordingly, adding another layer of competition:
That is, when a user has a task to accomplish, whether they are willing to hand that task over to your Agent first.