The Personal Agent battle spreads to hardware: the different paths of Qwen, Shokz, viaim and Bee

AI glasses, earbuds, and recording devices are becoming the entry points through which Personal Agents enter the real world.

On October 13, Qwen released AI glasses and AI clip-on earbuds, hoping to let the same Personal Agent understand users and carry tasks across different devices. At the same time, hardware makers are also trying to use earbuds and recording devices to turn real-world conversations into context that Agents can call upon.

One of the characteristics of standalone hardware is that it starts up quickly. In efficiency scenarios, the process of unlocking a phone, opening an app, and pressing the record button is somewhat long, and lighting up the phone to check recording status during a meeting can also be distracting. By contrast, pressing a button on a recording card and seeing the indicator light come on is simpler and more direct, making the user experience more relaxed.

Although model inference and task execution capabilities can be provided by external platforms, hardware makers can also participate in context management.

Qwen integrates models, apps, and hardware; Shokz relies on external AI capabilities to provide services; viaim, a brand under Future Intelligence, tries to use its own Personal Agent to orchestrate other Agents; and Amazon's Bee connects to existing assistant and shopping services. Different companies are exploring their own positions.

Hardware serves as the senses through which Personal Agents absorb context in the real world, and competition among vendors is unfolding around the long-term management of this information, task invocation, and service charges.

One

By the ear, before the eyes: the same Agent

A single Personal Agent tied together an entire hardware launch event.

On October 13, Alibaba's Qwen released the AI glasses N1 and N1 Pro and AI clip-on earbuds, with all three products going on sale in stock that day. After stacking national subsidies and limited-time discounts, the glasses start at 2,218 yuan, and the earbuds start at 1,099 yuan for the first sale.

Song Gang, vice president of Alibaba Group's Qwen business unit, said at the launch event:

"Qwen's goal is to become everyone's Personal Agent. We believe AI wearables are Qwen's extension in the physical world and an important part of the Personal Agent. No matter where you initiate a request from, behind it is the same Qwen, understanding the same person and continuing the same goal."

The glasses, earbuds, and recording card target scenarios such as visual perception, on-the-go voice interaction, and office recording, respectively. The "Skills Plaza" of the Qwen open platform has begun introducing third-party services, so that users' voice commands can invoke more functions.

In August, Shokz launched the OpenFit 2 AI earbuds; in September, Meta announced at the Connect conference that it would bring the personal agent Muse into its AI glasses product line, with rollout expected in a few months.

The AI-native story was already told two years ago by the Humane AI Pin and Rabbit R1, but the product experience did not go that far, and the novelty in the promotional videos was not delivered. Most of the new wave of AI hardware no longer talks about replacing phones; earbuds, glasses, and recording devices are meant to take on more perception and interaction tasks.

Having a single piece of hardware cover all of a user's life scenarios may not be the most reasonable choice.

Hardware such as earbuds, glasses, and recording devices has limited size and needs to be worn close to the body or carried around. If it comes down to feature integration and hardware specifications, it is hard to compete with smartphones.

The key is how to find each one's ecological niche, coordinate with one another, cover different usage scenarios and user needs separately, and ultimately serve a complete Agent workflow.

For multiple pieces of hardware connected to the same Agent, performance and specifications still matter and are part of the long-term user experience. On that premise, the hardware also needs to continuously acquire scenario information useful to the Agent.

Two

Hearing clearly is only the premise

For an Agent to use real-world conversations to complete subsequent tasks, it first needs clear, complete recordings. If the recording obtained at the outset is incomplete, it is difficult for subsequent models to accurately reconstruct the missing content.

The context an Agent needs during processing is not limited to text content. Historical recordings, transcriptions, documents, people, project background, and unfinished tasks may all be context needed to complete the next task.

Yan Gengzhen, China director of acoustic technology company Soundskrit, explained to All-Weather Tech that different forms of AI hardware have different requirements for microphones and sound processing algorithms.

Glasses first need to extract the wearer's own voice and remove surrounding noise in noisy environments; if used for face-to-face translation, they also need to hear the conversation partner clearly. Recording cards place more emphasis on long-distance pickup, complex meeting-room environments, and distinguishing between different speakers.

When All-Weather Tech asked which types of companies the new acoustic demands mainly come from, Yan Gengzhen replied:

The placement of the pickup device also affects context.

Earbuds worn by the ear are closer to the wearer and more easily prioritize clearly capturing the user's own voice in noisy public scenarios. But when holding meetings in the office, people often place the recording card on the table to ensure everyone's voice can be recorded.

Some products try to let both the earbuds and the charging case record. Ma Xiao, founder of Future Intelligence, introduced that the AI agent earbuds viaim Rise, released at IFA in Berlin in September, merged multiple modes such as online recording, face-to-face recording, and audio/video recording into a single recording button.

The button on the charging case can also start recording after the phone app is closed, so the recording function does not fully depend on the earbuds being worn and on the phone app.

"We also put a lot of effort into this box. There may have been a lot of opposition at the beginning, but we still bravely pushed it forward, because Agent hardware definitely needs to incorporate some new things."

Ma Xiao does not believe that this wave of Personal Agent enthusiasm has brought about a fundamental change in hardware technology. In his view, what has changed is the perspective from which people view hardware.

Agents connect devices with context, making a single recording the starting point for subsequent tasks. Ma Xiao believes that when Agents can drive meeting items to form a closed loop, the usage scenarios and commercial value of hardware will also change accordingly.

After sound is captured, it still needs to go through corresponding signal processing and model parsing before it can become information callable by the Agent. As for how on-device and cloud processing divide the work, it varies from product to product.

The Qwen Office Recording Card QwenNote A2 has a built-in 4G IoT card and can transmit sound to the cloud without relying on a phone for real-time transcription and meeting-notes processing. The original audio is deleted after transcription is completed, retaining only the text and meeting notes.

What viaim values is organizing multiple recordings, documents, and historical discussions into the same project for the Agent to call in subsequent tasks. The device side mainly handles tasks such as sound pickup, noise reduction, and transmission. When the network signal is poor, recordings are first cached locally and uploaded after the connection is restored.

Currently, what is uploaded to the cloud for processing is the original audio. Transcription and subsequent Agent analysis are mainly completed in the cloud, and users can export the original audio and transcribed text.

Three

From listening to remembering

Recording, transcription, context management, and subsequent Agent services can be provided by different companies.

Take the Shokz OpenFit 2 AI earbuds as an example. Shokz controls the hardware and the user interaction entry point, recording transcription and meeting-note generation are handed to Qwen, and Guangfan provides support for the relatively independent "AI Lab" function.

Shokz replied to All-Weather Technology that after the user presses the recording button, the audio is first stored in the earbuds' built-in chip, with each earbud supporting 4 hours of recording; after recording ends and the earbuds connect to the phone, the audio is automatically uploaded to the Shokz App, and then the Shokz server calls the Qwen large model for transcription and meeting-note generation.

At the Yunqi Conference in September, Shokz, together with Guangfan Technology, launched the "AI Lab" on the OpenFit 2 AI.

According to Shokz's reply, Guangfan provides technical support for functions in the AI Lab in the Shokz App such as intelligent dialogue, schedule to-dos, quick notes, and message assistant. The recording and meeting-notes service and the AI Lab service are independent of each other, and their data do not intersect.

Shokz also stated that its information security management system has passed Level 3 of the Cybersecurity Classified Protection and SOC 2 Type 1 certification, and that encryption measures are adopted for the process from recording and transcription to file archiving.

Qwen takes an integrated path, launching glasses, earbuds, and a recording card while also operating an App, model capabilities, and a third-party skills ecosystem, bringing multiple entry points into the same service system. If the same Qwen can understand users, retrieve historical context, and connect to Alibaba services, then Alibaba will have the opportunity to operate all the way from the entry point to task execution and service consumption.

The commercial value of this route lies in the closed loop; the challenge is that users may not be willing to hand all their daily tasks to the same ecosystem.

Plaud supports users in choosing different models. Users can initiate tasks through external AI tools such as ChatGPT and Claude, and call meeting materials saved by Plaud through connection methods such as MCP.

viaim hopes to use its own Agent to handle part of context management and task allocation. "We made an Agent that mobilizes other Agents," Ma Xiao, founder of Future Intelligence, explained to All-Weather Technology.

Ma Xiao introduced two ways in which viaim cooperates with Tencent WorkBuddy. One is that the host side schedules, and its own Personal Agent analyzes the information accumulated by the user, identifies needs, and hands the task to WorkBuddy for execution. The other is opening the CLI, where the user initiates a task in WorkBuddy and retrieves recording records through the viaim CLI.

The two modes can use the same batch of recordings; the difference lies in where the task is initiated.

On October 15 local time, Amazon-owned Bee released the Bee Frontier wristband and the magnetic card accessory Bee Eclipse.

Both hardware products can extract information from users' conversations and connected apps and generate summaries and to-dos. The paid subscription service Bee Pro supports third-party apps and integration with Alexa+ and Amazon.com.

Also making hardware, manufacturers have chosen different positions in the value chain. Some operate hardware and user entry points, some hope to control context and task allocation, and some platform companies try to connect real-world information to their own services. As for how often users actually call on this information and their willingness to pay, the data publicly disclosed so far cannot fully explain it.

Four

Context processing is a long chain, and different companies cover different links in this chain. In a specific task, what each company charges is not a fee at the same layer. Selling hardware is only the first step; in long-term use, software services and subscription systems are important sources of revenue.

Many manufacturers have already launched subscription services, but when it comes to specific pricing standards, each company is trying to design paid benefits separately for information processing volume, context management, and subsequent task services.

Over the past six months or so, Alibaba, iFlytek, StepFun, Tencent Hunyuan, Xiaomi, and others have successively released or upgraded speech recognition models.

When Qwen released the Qwen-Audio-3.1 series in September, prices across its entire speech model lineup were cut, with speech recognition dropping by as much as 95%. General-purpose devices such as phones can call speech recognition capabilities through software, while the subscription value of standalone hardware still needs to be demonstrated through context management and Agent services.

For users who frequently attend meetings or conduct interviews, if they have to check the remaining duration every time before deciding whether to record, the quota itself may become a barrier to use. But if the user base expands, the revenue per user may fail to cover the corresponding costs, and the gross margin of the subscription business may still come under pressure.

As users continue to use the hardware and Agent services, the number of recordings grows, but the usable context does not necessarily increase proportionally. Context management also needs to organize multiple interactions into project background, personal relationships, and historical decisions, so that the Agent can continuously draw on them in subsequent tasks and improve execution efficiency and delivery results.

For example, after a meeting ends, the Agent needs to accurately identify the person responsible and the deadline, and sync the to-dos to the user's office system.

In this way, when the user discusses the same project again, they can easily retrieve the previous task results and to-dos. If the Agent needs to send messages to external parties, it should also request user confirmation at the necessary steps.

In the past, hardware product sales were usually one-off transactions, and upgrades often meant buying another device.

Personal Agent tells a long-term story, hardware is beginning to emphasize "always on," and context is becoming a resource—but a resource itself does not directly accumulate into a competitive moat, nor does it prove that making Personal Agent hardware is necessarily a good business. The tasks the Agent executes are becoming increasingly complex, which may bring both new revenue and new costs.

The long-term value of Agent hardware depends first on product capability. Few people are willing to pay for an Agent assistant that frequently loses its memory and says whatever comes to mind, and proactively providing services also requires a sense of propriety.

Once product capability is in place, user demand must be considered. If users leave the hardware idle most of the time after purchase, or resell it once the novelty wears off, and ultimately call Agent functions infrequently, then the main function the hardware serves is still collection rather than accumulation.

On the business model side, the composition of current subscription plans varies, and manufacturers have different considerations regarding the value brought by Agents. For the subscription business to keep running, users need to obtain perceivable long-term value, and manufacturers also need revenue to cover storage, inference, and task execution costs.

This article is from the WeChat public account "All-Weather Technology" (ID: iawtmt), author: Lü Zhetong, published by 36Kr with authorization.