China News Weekly Reporter: Zhou You
Published in the 1259th issue of "China News Weekly" magazine on November 9, 2026
Recently, Anthropic released a mobile application called Claude Duet. It will push a briefing to the user every morning, listing the most important things in the mailbox, calendar and each account; when it finds that a bill has not been paid, it will draft a reminder email; with the user's nod, it can also directly place an order for payment. And this is just one of the same products that global technology giants have launched intensively in the past month or so.
In September, Meta took the lead in launching the personal intelligence Muse. Immediately afterwards, Microsoft released a new generation of intelligent robot Copilot, calling it "an assistant that is still working when you are not at work." At the end of September, OpenAI launched the product dots, which can send messages and make calls to others. At the end of October, Tencent WorkBuddy released Buddy One, a collaborative system intelligence that connects documents, meetings and network disks.
These products have a common name: personal intelligence. When mobile applications actively report to their owners and complete tasks, AI is evolving from a chat tool to a work assistant, opening up a new working model for humans.
"Turn off the computer, it's still working"
Recently, a steamed stuffed bun shop owner in Changsha used Tencent WorkBuddy to set up an order workbench to automatically sort out the orders in the group, generate order calendars, and even hand over the signboards, menus and posters to the intelligent design. What used to be three people's work can now be done by one person and one machine, and the daily sales of steamed buns have increased from 300 to 800.
Judging from the capabilities of each company's products, personal intelligence has already done a lot of work. When Muse goes online, it will be able to compare prices, send and receive emails, manage calendars, and book trips for users. Claude Duet is aimed at daily official tasks. He can search for invoices in accounting software, draft reminder emails, and prepare contracts to be signed. The dots developed by OpenAI are good at active research. If the user does not look for it, it will also think about "what else can it do to help the owner" in the background. Copilot takes away documents, forms, and presentations, and can also keep an eye on meetings and organize to-dos for users.
“Proactively looking for people” has become a standard capability of personal intelligence. An ordinary intelligent agent is like a receptionist at a shopping mall information desk. It answers only one question when the user asks it, and forgets it quickly. A personal agent is more like a full-time housekeeper who lives at home. When the user gives it a task, it will break it down into steps and then report back to the user. More importantly, even if the user turns off the computer or mobile phone, it is still working.
According to Wu Shuang, NVIDIA's chief engineer for autonomous driving, two key conditions are needed to achieve the persistence of background tasks: First, continuous contextual memory, that is, sufficient model capabilities. Wu Shuang believes that the model capabilities have matured, so personal intelligence will explode this year.
The second is an environment that allows intelligent agents to always exist and work. This is also the reason why everyone puts their intelligence into the cloud. "In the past, everyone thought that emails had to be stored in their own computers, but later they gradually became accustomed to using web mailboxes and the cloud." Wu Shuang said that only in the cloud can the intelligent agent work 24 hours a day. This is equivalent to giving each agent a dedicated office in the cloud.
The way the personal agent works is reminiscent of the "lobster", also known as OpenClaw. It can also manage tasks in the background for long periods of time. Meta has publicly stated that it is deeply inspired by OpenClaw in its product concept, but the two are fundamentally different. OpenClaw is not an independent large model, but a self-hosted agent framework. Users need to deploy, access large models, and configure tools by themselves. It is indeed "hard-working", but its "brain" is the model selected by the user, and its performance depends on the strength of the connected model.
Personal agent users do not need to perform any configuration. According to Meta, it allocates an independent cloud virtual machine to each user and sets up an independent agent to approve external operations. Users can view audit logs. For example, sensitive operations such as payment require user confirmation. In other words, OpenClaw lets users take charge of their own security, while Muse's security is taken over by Meta.
The technical threshold determines that OpenClaw is aimed at programmers and developers with technical capabilities. Wu Shuang pointed out that personal intelligence is aimed at ordinary consumers.
Many ordinary people have begun to try personal intelligence. According to official data from Tencent, WorkBuddy has completed a total of 9 billion task requests since its launch. Apptopia, an American market research company, released data in October. Within 12 days after Muse was launched, the number of iPhone downloads in the United States and Canada reached 1.8 million, exceeding the data of the ChatGPT mobile version during the same period when it was launched, and it once topped the U.S. app store list.
However, the experience of personal intelligence cannot be called perfect. A Dots user told China News Weekly that when Dots makes voice calls, it sometimes takes several minutes of repeated attempts to connect, and there are also lags when taking over computer operations. Anthropic recently admitted that Claude Duet is not yet good at complex personal matters such as booking tickets, and may still misread the news and misjudge the priorities of events.
Muse is currently free to use. Meta stated at a developer conference on September 23 that in the future, it plans to extract a small handling fee from each transaction completed by the agent instead of charging a subscription fee. OpenAI dots are included in premium subscriptions starting at $100 per month, and Claude Duet is included in the $20 per month package.
crisis of confidence
In October, Wal-Mart announced that it would open a special shopping channel for smartphones, create a separate profile for each smartphone, and set a consumption limit, which is equivalent to applying for a limit supplementary card to the user's credit card, and the user can withdraw the authorization at any time. However, more platforms are not willing to open their arms to intelligent agents.
At the end of September, after Amazon discovered that Muse could access its platform, it asked Meta to actively exclude Amazon from Muse's support, but Meta rejected this request. Subsequently, Amazon began to block Muse. When users try to shop on the Amazon platform through Muse, they will receive a prompt stating that "continuous access by unauthorized AI agents violates Amazon's terms of use." Amazon stated that third-party intelligent agents should operate openly when shopping for customers and respect merchants’ decisions to participate, but Muse failed to do this.
Muse has also staged the drama of "replacing real people with AI". At the beginning of the launch, many merchants hung up the phone as soon as they heard that the other party was AI, causing Muse's core functions such as reservations and ticket bookings to be blocked. To get around this hurdle, Meta launched an internal test called "human concierge." When an AI call fails, Muse can transfer the request to a trained human contractor, who will make the call and complete the task. The test feature was quickly withdrawn because users could unknowingly expose sensitive information to contractors.
Zhou Zhou, vice president of vivo and director of the AI Global Research Institute, recently told the media that the mainstream mode of intelligent work in the future will be a protocol, which is to establish a dedicated line between the intelligent agent and the merchant. Compared with simulating human hands to click and operate the user's mobile phone to make calls, direct connection is more decent and efficient.
Even so, an AI that can send emails and pay money needs to know too much information, and personal information also poses security risks. According to a report in the US "Wired" magazine on October 3, security researchers discovered from Muse's internal documents that it updates profiles for relatives and friends in users' lives hourly, including preferences, experiences, and relationship statuses. By mid-October, another vulnerability was exposed in one of Muse’s conversation sharing functions. Third parties can use the conversation history to see corporate loan applications, internal revenue reviews written to the CEO, and even bank fraud alerts. Meta subsequently suspended the feature.
The risk of fraud is also unavoidable. There is an attack called "cue word injection", which is disguised as an instruction to the AI, just like a scammer inserting a fake "master note" into the butler's inbox to induce it to do things that the master did not tell it to do. Therefore, the security of personal intelligence requires layers of defense. Claude Duet will show users exactly what they need to do before sending messages, making payments, or signing documents. Buying and selling securities and account transfers are prohibited, and even if the user instructs the user to do so, the agent will not do it.
American technology media The Verge published an analysis in October saying that protecting user privacy is becoming a selling point that various companies are competing to sell. The analysis also pointed out that Muse allows Meta to use user data to train models by default, while dots needs to gain control of the entire cloud disk. In other words, the performance of many intelligent agents is itself based on the user's personal information.
Currently, Muse is still only open to users in the United States and Canada, and Claude Duet needs to queue for an invitation. The explosive growth of users in the early days of Muse's launch clearly subsided after about a month. Wu Shuang believes that if users feel that the rewards brought by personal intelligence, such as the convenience of handling affairs, are worth paying for personal information, they will naturally pay for it. The premise is that users trust the manufacturer's ability to protect personal information security.
"Many people still need to adapt to and get used to the daily use of intelligent agents. What personal intelligence now requires is product strategy, function polishing and engineering optimization." Wu Shuang said that the future of having a personal agent in everyone's hands is almost a certainty. Before actually entering daily life, these AI butlers must first convince users and the applications on their mobile phones.
"China News Weekly" Issue 41, 2026
