Tencent Marvis's distinct approach to building Agents: taking users' computers under "management"

By Li Zhaofeng   

Edited by Zhang Yuxin

Recently, Tencent launched a new version of Marvis, upgrading the positioning of this personal Agent with the "dark horse" as its image to an "AI butler". According to Smart Emergence, as of this month, the number of daily active users of Marvis has increased nearly fourfold compared with its launch in May.

Marvis is developed by Tencent's Yingyongbao team, which focuses on app distribution. When it was launched in May this year, it was called an "operating system-level personal AI assistant", with the focus on completing tasks such as installing software and finding files according to user instructions.

In this revision, in addition to computer file management, Tencent has placed computer system and hardware management in a more prominent position: Marvis can check battery health, identify high-power-consuming apps, and adjust startup items. Cai Jiantao, head of the Marvis business, said at a recent event that since its launch in May, users' actual usage has been different from the team's original vision, "Many users really use us as an AI butler."

Previously, Cai Jiantao disclosed in a public interview that local files are the most frequently used scenario, involving 44% of users; another 28% use it to manage computers and hardware. Feedback from some early Marvis users shows that after downloading Marvis, they did not know what tasks to assign to it. In response, Cai Jiantao said, "Half of this revision is designed to solve this problem." The new version focuses on four types of tasks: computers, files, software and browsers.

At present, the scenarios covered by the large number of Agents on the market have been clearly segmented.

The earlier segment targets office scenarios. In March, Anthropic opened a research preview for Claude Cowork to operate computers: with permission, it can use desktop applications to complete briefings and reports. In China, office-focused Agents such as WorkBuddy, Qianwen Office and Doubao Work also allow users to assign work tasks and retrieve outputs such as documents, web pages and PPTs.

Another segment targets online life scenarios. Meta launched Muse this month, allowing it to process emails and travel bookings in a cloud virtual computer; Alibaba's Qianwen App also announced its entry into the personal Agent track, integrating multiple life scenarios such as food ordering, shopping, and flight and hotel bookings. Global manufacturers have successively launched personal Agent products — according to a survey of 5,067 American adults conducted by U.S. venture capital firm Menlo Ventures and market research firm Morning Consult in July this year, 41% of respondents who use AI have tried AI Agents, and 24% use them regularly. At least among AI users in the United States, Agents have moved beyond the stage of being a niche novelty for early adopters.

Office Agents and online life Agents need to rely on PC platforms to complete tasks, while what Marvis wants to do is manage the user's terminal itself, and act as a "butler" for all scattered tasks on the PC side.

In addition, Marvis has a relatively unique Agent invocation path: it first uses the end-side model to build an index for authorized data, then uses the index for search when users look up content, and invokes the cloud-side model on demand — this is quite different from the idea of office Agents such as WorkBuddy, which basically rely on cloud-side models.

01 What progress has Marvis made as a "butler"

To realize "managing the computer", it is necessary to go deep into the system layer first.

In the new version demo, Marvis adjusted the startup items of a Mac, and identified high-power-consuming applications that affect battery life. However, when encountering operations that require system permissions, it will stop at the confirmation window and wait for the user to enter the password. How much it can help users still depends on the functions opened by the operating system.

Similar operations can be seen in similar products. Microsoft added its own Agent to the settings of some Windows computers last year, and later tried to open file and setting functions to third-party Agents, on the premise that users must actively enable relevant functions and agree to access sensitive information.

The new version of the file butler function adds a library update reminder. In the public demo, the Marvis team took a campus recruitment scenario as an example: after the HR built a resume library, 50 more resumes were received. When the new files are saved to the computer, Marvis will prompt to add them to the existing library, but this process ultimately requires user confirmation.

Cai Jiantao said, "When Marvis searches for local files, it will use the end-side model to process file content and build an index within the scope authorized by the user, which can greatly speed up the file search. Some Agents do not have the step of 'building a file index with the end-side model', and if you ask them to find a file from five years ago, it may take an hour."

Local processing requires computing power, which was once a device threshold for Marvis. Cai Jiantao said that the May version required a six-core processor and 16GB of memory, mainly because the knowledge base consumes a lot of computing power; the new version simplifies the construction process of the local knowledge base, which reduces the required computer configuration to a four-core processor and 8GB of memory, allowing users to "get started easily".

Marvis also plans to expand file search to private storage devices such as NAS (private cloud/network attached storage). At present, it has built a customized NAS in cooperation with Intel and Feiniu, and plans to launch it in October. In the previous demo, the capability of Marvis combined with NAS has been preliminarily demonstrated: when users search for "sea", Marvis can locate the clips where the sea surface appears in the video; when searching for "videos with applause in the sound", it can quickly locate the corresponding audio clips.

A Tencent insider explained the process behind this video search: within the scope authorized by the user, Marvis first decodes the video, processes the images and sounds, and then builds an index. For a NAS with more than ten terabytes of data, it will first process the materials that the user has paid attention to recently, and then expand the scope according to search demands.

In response to users' demand for installing software, Cai Jiantao said that about 40,000 to 50,000 people use it to do this every day before the revision. The new version centralizes demands such as software installation, update and operation into the independent function of "Software Butler", making it easier for users to find.

At the same time, the Marvis team also demonstrated its application in gaming scenarios, and plans to develop it into an independent "Game Butler" in the future. In the Steam game skill demo, Marvis will analyze the battle after watching a round of game; in the teaching mode, it will prompt players which key to press next during the game.

In terms of operation entry, users can remotely send instructions to the computer through their mobile phones. Cai Jiantao said that since the launch of Marvis, remote control has been used more than 18 million times in total.

According to Smart Emergence, Tencent is also seeking cooperation with computer manufacturers to bring Marvis to more devices. At present, the OEM version for complete machine manufacturers has been launched.

At present, the AI assistant market on the PC side has entered a stage of fierce competition. Some manufacturers are adopting self-developed routes. For example, Lenovo's Qira has been opened for download by eligible computer users; HP's AI Companion has been pre-installed on some Copilot+ PCs.

"Some PC manufacturers have their own AI assistants, while some do not want this type of product form." Talking about cooperation with PC manufacturers, Cai Jiantao said that Tencent can also only provide the underlying AI engine, and the manufacturers decide what the product form for end users will be.

Talking about computer diagnosis, Lin Songtao, Vice President of Tencent, said in a previous media interview: "What is lacking is never knowledge, but the understanding of this specific device." Cai Jiantao envisages that if computer manufacturers open more device status data, when the C drive is almost full, Marvis can actively remind users to clean up, without waiting for users to send instructions first.

02 Unavoidable office scenarios?

Marvis does not focus on office scenarios, but as long as it starts managing users' files, office scenarios are unavoidable.

"From the first month of our project establishment, we decided not to position ourselves as an office AI." Cai Jiantao said when talking about the product direction earlier. Documents and PPTs are the first tasks chosen by many Agents: the output is easy to see, and it is "much less difficult than making a game assistant". Marvis will also meet office demands, but "we rarely say we are the best Agent for making PPTs, because it is relatively far from our main line".

Within Tencent, WorkBuddy is the product that focuses on office scenarios. According to statistics from Analysys, in June this year, the total visits of 17 monitored China desktop AI-native office Agents exceeded 60 million times, of which WorkBuddy reached 20.97 million times, accounting for one third.

Not long ago, Tencent has begun to integrate a large number of Agent products. Just yesterday, Tencent announced that QClaw, which was recently merged into the WorkBuddy team, will cease operations on December 24, and users can migrate their data to WorkBuddy.

Despite the positioning difference, it is difficult for Marvis and WorkBuddy to achieve complete differentiation in product functions — WorkBuddy can read and write local files authorized by users; Marvis can also write reports.

However, compared with Marvis, which is still exploring its positioning, WorkBuddy has a clear target market and has made progress in commercialization. James Mitchell, Chief Strategy Officer of Tencent, said at the August earnings call that user payments for WorkBuddy are growing significantly, and the gross margin of paying users is comparable to the overall level of Tencent Cloud.

Talking about the commercialization of Marvis, Cai Jiantao took the C drive cleanup tools on the market as an example: developers spend years sorting out file types, directories and installation caches, so that they can clean up more thoroughly and faster than other tools. "It is reasonable to pay for this special capability."

He said that Marvis will launch a commercialization plan when the product has sufficient competitiveness; the charging point "may not come from the basic capabilities of the Agent itself", but allow users to pay for ecological partners that provide professional tools.