Article | Sleepy, Sīwéi Guàiguài
In December 2025, the long-rumored 'Doubao Phone' finally debuted. It installed the Doubao Phone Assistant technical preview into a Nubia M153 engineering prototype, launched at a price of 3,499 RMB. The first batch of approximately 30,000 units sold out on the day of release.
We remember that in the few days just after launch, its price on second-hand markets soared, commanding a premium many times over the original price. The Beating editorial team even bought two.
It wasn't because it was a particularly good phone. On the contrary, as a 'technical preview,' the first-generation Doubao Phone's experience wasn't great. What excited us was that it was the first time AI was pulled out of the chatbox, transformed from a Chat bot into an AI Agent capable of controlling a smartphone.
On the Doubao Phone, the AI could see the screen, understand what you were browsing, hear you speak, jump between different apps, and directly help you do many things, like checking train tickets, shopping for price comparisons, claiming coupons and placing orders, or editing photos. Although sensitive steps like payment still required user confirmation, it could indeed independently complete many operations that in the past required us to tap through step by step.
While it was still a bit clumsy—sometimes responding slowly, sometimes freezing up, awkward like someone just learning to use a smartphone—it truly gave us our first intuitive feel for how convenient AI could be in daily life.
Later, Lobster emerged and took the world by storm. AI Agents became another 'iPhone moment' in the AI field after ChatGPT's debut. A slew of manufacturers and entrepreneurs began selling computers and phones pre-installed with OpenClaw. The Doubao Phone was at least a version ahead of them. You could even say the Doubao Phone was a pioneer in this wave of Agent enthusiasm.
But unfortunately, the Doubao Phone soon ran into a blockade by the major platforms. Scenarios like WeChat, Taobao, Alipay, and banking apps successively encountered access or operational restrictions. Some called it a 'block,' others said it triggered risk controls. But for users, it made no difference—it simply stopped working.
We felt great regret. The Doubao Phone was certainly not a mature consumer electronics product, but it showed the entire industry the prototype of the next-generation entry point.
So even though the hype around the Doubao Phone has faded, we never completely let the matter go. Until recently, our routine information scraping captured thousands of job postings. Analysis revealed that ByteDance seems to be reviving smartphone research and development.
Three Categories, One Clue
We scraped three categories from ByteDance's official social recruitment page: AI Innovation Business, Mobile OS, and Doubao Phone Assistant.
After deduplication by position ID, we further scraped detailed page information and cross-referenced keywords from job titles, job descriptions, and qualification requirements for organization.
Unlike recruitment for ordinary AI app teams, ByteDance's batch of social recruitment positions also includes roles related to mobile phone systems, cameras, touch controls, connectivity, battery life, heat management, chip adaptation, structural design, device assembly processes, and production line testing.
These terms are uncommon within internet companies; they are the daily concerns of smartphone manufacturers, supply chain companies, and engineering teams.
ByteDance is hiring people to go into factories.
This doesn't definitively confirm ByteDance will launch its own phone brand, but it at least indicates they are restarting R&D work on phone-level hardware.
Let's look at what these positions themselves reveal.
Doubao Phone Assistant: From Answering Questions to Performing Tasks
First, the Doubao Phone Assistant.
We performed a more focused filtering on the original data, searching for positions where the name, description, or requirements mentioned 'Doubao Phone Assistant'—a total of 83 positions. These could be divided into three broad categories, which together happen to outline the shape of a system-level AI Agent.
The first category of positions is responsible for enabling the AI with Agent capabilities.
For example, the 'Agent Development Engineer - Doubao Phone Assistant' position states the need for the AI to perform task decomposition, context organization, tool invocation, memory retrieval, state management, result verification, and exception recovery—the foundational capabilities of all AI Agents we use today.
The second category focuses on giving the AI Agent a good memory.
Positions mention directions like 'perception and memory,' 'user memory,' 'personal knowledge graph,' and 'long-term preferences.' For an AI Agent to truly integrate into our lives, it can't start from scratch every day; it needs reliable, stable long-term memory.
Of course, this easily touches on issues of privacy and boundaries. But judging from the recruitment materials, ByteDance has at least begun treating 'memory' as one of the most important capabilities of the Doubao Phone Assistant for R&D.
The third category of positions is responsible for enabling the AI Agent to utilize those capabilities within a phone.
If the Doubao Phone Assistant is to operate the phone for the user, it can't live only in the cloud, nor can it be just an app. It needs a complete set of capabilities—including models, memory, task execution, on-device deployment, system applications, audio/video, communication, testing, and quality assurance—to understand the user's speech, comprehend the environment, collaborate across devices, stay on standby, and avoid causing problems.
Mobile OS: The Phone's Underlying System is the AI Agent's Hurdle
Now, look at Mobile OS.
There are 236 positions related to Mobile OS, with primary office locations concentrated in Beijing, Shanghai, and Shenzhen. Terms repeatedly appearing in the job descriptions are kernel, chip, drivers, camera, display, audio, network, power consumption, thermal management, mass production delivery. These are almost all terms closer to hardware and the phone's underlying system.
For example. The responsibilities for the 'Kernel Leader - Mobile OS' position state the need to lead the memory and storage team in kernel adaptation and development for new Qualcomm platforms, ensuring the system works well with mainstream phone chips, and managing the phone's memory and storage. These capabilities are key for an AI Agent to achieve real-time responsiveness and handle tasks in the background.
Furthermore, positions mention terms like SoC, BSP, and RTOS. SoC can be roughly understood as the phone's core chip, BSP is the underlying software that lets the system and hardware recognize and cooperate with each other, and RTOS is often used in scenarios with high demands for responsiveness and power efficiency.
So the signal from the Mobile OS positions is that ByteDance is hiring people who understand phone-level device systems. They need to know at least where an AI Agent running on a phone might get stuck by permission issues, power consumption problems, or system stability issues, and which problems require joint solutions with chipmakers, manufacturers, and testing teams.
Judging from the requirements of these active job postings, ByteDance is already entering the deep waters of smartphone development.
Shenzhen Coordinates: Signals of Hardware and Mass Production
It's also necessary to look separately at positions located in Shenzhen.
If positions in Beijing lean more towards models, algorithms, and platforms, and those in Shanghai towards products and engineering, then positions in Shenzhen are often related to hardware, supply chain, testing, and mass production.
For a project that's just a cloud service, Shenzhen isn't that crucial; once it involves physical products, Shenzhen becomes very important.
And what we see in the Shenzhen-related positions are precisely these things.
Some positions are for human-computer interaction design, covering hardware physical interaction, software interface interaction, and multi-device linkage experience. These positions don't just consider how to design the on-screen interfaces; they also consider the feel of the physical device, buttons, how to wake it up, and how to link with other devices.
Other positions are closer to the engineering floor: interconnection, power consumption, short-range communication, baseband, device assembly process, structure, and testing processes.
These terms don't sound as glamorous as 'agent,' 'multimodal,' or 'world model.' But in consumer electronics, these are ultimately what determine life or death.
If ByteDance only wanted to make Doubao a better phone app, it wouldn't need to do all this hard work. Once it starts hiring for these positions, it indicates readiness to board this ship.
ByteDance Can't Just Make Apps
In the past, the phone was a container for apps; in the AI era, the phone might become the body of the Agent.
If the phone is just an app container, then a company like ByteDance can build its kingdom through individual apps based on content, algorithms, and product prowess. But if the phone becomes the Agent's body, and the user issues a task first, whoever can take that task has the chance to decide the subsequent path.
In this path, apps are demoted to callable tools. This makes all super apps uncomfortable. Because Agents inherently bypass the middle layer.
So, the real difficulty may not lie in whether Doubao can open an app, but in whether others are willing to let it open. And an AI that makes decisions for users cannot be granted passage as easily as a regular app.
For an Agent to move from the chatbox to the operational layer, it must handle a heap of dirty work that didn't belong to AI teams in the past. They need to know when the system kills background processes, when an operation triggers risk control, why the phone heats up, why factory yields aren't improving. In the past, AI teams didn't handle these things, but now they're unavoidable.
That's why ByteDance is hiring for these positions. It might not necessarily launch its own phone, but ByteDance certainly can no longer afford to be just an app inside someone else's phone.
For large model companies to become the next generation's user entry point, they cannot forever reside within someone else's operating system.
