Author | Cheng Yanjing, Editor | Hu Min
On overseas social media, an entrepreneurship blogger used the term "Agent Sprawl" to describe a phenomenon:
The more Agents there are, the more fragmented their permissions, data, and invocation relationships become. Enterprise organizations are also facing a new problem: Who should manage these AIs?
Similar issues have already emerged in consumer scenarios.
In the AI era, files, applications, browser pages, and data on users' computers are becoming a form of personal digital assets.
IDC predicts that by 2026, the number of active Agents globally will reach 79.4 million, nearly tripling compared to the previous year.
Agents are comprehensively penetrating the PC end, but the industry rarely focuses on the subsequent issues: Who should clean up the system redundancies left behind after they complete their tasks?
For example, AI-generated files crowd the hard drive, the C drive suddenly turns red, and background processes multiply. Users originally introduced Agents to improve efficiency, but ended up spending more effort cleaning up the environment.
Just like the early PC butlers that could help users solve problems such as antivirus, software installation, and uninstallation, this role is surprisingly absent in the AI era.
The vacancy itself is an opportunity for a new category. This is the question the "AI Butler" needs to answer, and also the reason why Marvis, Tencent's OS-level personal AI butler, announced a positioning upgrade four months after its launch.
01. From "Finding Apps" to "Finding Services"
In the era of traditional software, human needs usually arise first, followed by searching for the corresponding tools.
For example, opening Word to write documents, Photoshop to process images, and a browser to search for information.
But with the advent of AI, it is time for these things to "work" on their own.
When a user says, "Help me organize a market analysis," they do not necessarily need to know which software to open, where the materials are, or which menu hides the specific functions.
After understanding the goal, the AI can find information, invoke tools, and process files on its own, and then hand back the results.
The interaction between users and computers has thus changed: in the past, it was about finding apps; now, it is increasingly like finding services.
In early September this year, the head of the Marvis ecosystem proposed the concept of "Direct AI Services" at the Arm conference. The logic behind it is not complicated: once Agents truly enter the workflow, the standard for measuring an Agent is no longer just how many tools it invokes or how many capabilities it has, but whether it can truly accomplish the tasks assigned by users.
Under this logic, the applications themselves retreat to the background, becoming callable execution nodes; developers also shift to providing services to AI through methods such as MCP and SDK.
This is also consistent with the view of Lin Songtao, Vice President of Tencent, during this year's WAIC: AI applications are experiencing a turning point "from answering all questions to fulfilling missions," and "without delivering results, there is no way to form a closed-loop business."
However, the more capable AI becomes at working, another problem arises.
After a task is completed, what is left on the computer includes: AI-generated files, images, and code, browser information, downloaded materials, newly installed software, accounts, and various local data.
Essentially, these are all users' digital assets, which need to be managed: they need to be tidy, secure, and findable when invoked. But this is precisely the blind spot that task-based billing Agents cannot cover.
This naturally defines two types of AI roles: the Agent is the "helper," and the AI butler is the "manager."
The helper is responsible for "understanding, executing, and delivering," settling by the task, and leaving once done, with no obligation to remember or maintain the computer's system environment.
The "butler" is another role: it understands your computer, files, devices, and digital assets over the long term, and takes action on your behalf when needed.
Just like Marvis's "butler" positioning: Marvis is like establishing a set of "house rules" or order for the user's "digital home." Its service boundary is not to complete a single task, but to manage personal digital assets over the long term.
This is a clear division of labor taking shape: Agents are responsible for working on the front end, while the butler is in charge of overall management on the back end.
The former is already a red ocean, while the position of "device butler" in the AI era is currently almost a blank slate. This also makes the next industry proposition extremely specific: What exactly should this butler manage, and how?
02. How the Butler Manages: Experience as the Foundation, Boundaries as the Limit
Sometimes when encountering a slow computer boot, you either endure it or follow tutorials to manually troubleshoot one by one in the system settings.
Now you just need to say: "The boot is too slow, help me deal with it."
Within a few minutes, startup item optimization and high-energy-consuming application troubleshooting are automatically completed. More importantly, there is a sense of proportion in the operations: for modifications involving underlying system permissions, the system will forcefully pop up a window requiring the user to enter a password for confirmation, never bypassing security mechanisms.
Accomplishing this is Marvis, Tencent's personal AI butler.
The new version solves a common industry problem: users often do not know what to talk to the AI about. Marvis directly materializes its capabilities, launching a "Butler Team" in the sidebar. The PC butler manages performance faults, the file butler manages the knowledge base, the software butler manages applications, and the browser butler is responsible for web automation tasks (game and life butlers will also be launched later).
Let's first look at the PC butler, which is most closely related to users.
The difficulty in fixing a computer has never been in "execution," but in "attribution": for example, when reporting a repair, a user will not say "DLL dependency corrupted," but rather "Tencent Meeting prompts me that a dll file is missing, help me fix it"; they will not say "the monitor refresh rate dropped," but "dragging windows is not smooth."
How to translate colloquial language into professional terminology is a technical job.
Part of Marvis's translation layer comes from Tencent's Enterprise IT Department—internally codenamed "8000"—a professional maintenance team that has served tens of thousands of Tencent employees and handled a massive amount of complex software and hardware issues over the past decade.
The core of the cooperation between the two sides is by no means a simple "integration of experience," but the engineering precipitation of the "maintenance feel" in the minds of IT experts into a "real case library" and "diagnostic paths" that can be invoked by Agents.
Opening the Skill list co-built by Marvis and the 8000 team, you can clearly see this "difference in judgment."
This list covers four major categories: operating systems, hardware, software, and networks, encompassing 27 real O&M scenarios. It deals not with standardized system commands, but with highly ambiguous user colloquialisms:
For example, when a user complains that "the colors look washed out after turning on High Dynamic Range (HDR)" or "the computer fan is very loud and hot, but I only have a few normal programs open," a general model would provide a troubleshooting encyclopedia from a search engine for the user to try themselves; whereas Marvis follows the real maintenance tree of the IT team—querying cross-data of the battery and high-energy-consuming processes, troubleshooting conflicts between graphics card drivers and color configurations, and locating the underlying environment variables of development tools.
Among them, hardware and peripherals (such as no signal on dual-screen monitors, screen tearing and color differences, Bluetooth keyboard connection failures) are the most heavily covered parts of this list. When encountering these issues, Marvis will step by step retrieve system logs, compare the fault characteristics of historical real cases, ultimately pinpoint the crux of the problem, and provide a repair solution.
This is the value of experience precipitation: general models only know "how," but an AI butler with diagnostic paths knows "why" and "where to intervene most appropriately."
A more typical example comes from an organic "good review" by a real user: unfamiliar programs appeared in the boot startup items. Marvis checked the installation logs to restore the scene—it turned out that another software brought it in by default as a dependency during installation.
The file butler illustrates another issue: managing the house and doing the work should never be separated.
Knowledge bases on the market require users to categorize and upload files themselves. Marvis does the opposite, changing it to "files stay in place, and the system automatically clusters them." Local paths remain unchanged, and user habits are not altered. When a large number of new files (such as batches of resumes) enter, the system will prompt to merge them with one click.
It not only integrates with Tencent Docs (WeCom will be integrated later), supporting multi-person co-creation and Q&A in the cloud, but also plans to use the subsequently launched "Project" capability to consolidate multiple tasks and deliverables within the same context.
This confirms a core logic: data management should not be the "sweeping monk" (the one who cleans up after the work is done), but a part of the process itself.
The software butler, on the other hand, moves from "managing applications" to "managing services." On the gaming side, Marvis has collaborated with Microsoft, Intel, and Qualcomm on tuning at the system and hardware levels. In the demonstration, the Steam AI tutorial provides combat analysis after a round, and the teaching mode prompts key presses in real time.
The game butler will be launched later as an independent role. On the life services side, the Luckin Coffee co-branded event involved over 200,000 participants and achieved over 10 million exposures across the network, with more than 50 service providers integrating—ordering takeout, drinking milk tea, and buying movie tickets in Marvis are gradually becoming daily AI capabilities.
It can be clearly seen in the actual demonstration that whether it is optimizing redundant boot startup items or repairing drivers involving the registry level, as long as core system permissions are touched, Marvis never bypasses security mechanisms, but forcefully pops up a confirmation box, requiring the user to be informed and enter a password.
This kind of "perfectly measured" diagnosis is precisely the quality of an excellent AI butler.
03. From One Computer to All Devices: The Layout Behind Marvis
The Marvis team shared an interesting daily routine: three computers, one pad, two mobile phones, and the camera's memory card.
But most of the time, he only talks to his mobile phone.
This kind of problem is very common in reality: users usually only express their needs through their mobile phones, but the actual files and computing power may be on another device.
Therefore, cross-device is not just "multi-device login." The mobile phone can be responsible for initiating requests, the PC and NAS for local execution, and the cloud for supplementing computing power when needed. The key is to enable a task to be continuously completed across different devices.
Taking a step further, what Marvis needs to solve is not just cross-device, but how to enter the system layer of more devices.
More importantly, for an Agent to truly complete a task, it must first know "what is in front of it."
Marvis incorporates local files, applications, and system status into the on-device index, allowing the Agent to already have a part of the device context when receiving a task, rather than searching from the cloud every time. For tasks involving a large amount of local data, this system-level capability is difficult to obtain solely by replacing a base model.
This is also the reason why Marvis needs to "join hands" with the hardware ecosystem.
For manufacturers of PCs, NAS, etc., building a complete set of Agent capabilities requires long-term investment in model, engineering, and product teams; if they directly access a mature Agent engine, they can add new AI capabilities on top of existing hardware and systems.
For Marvis, entering OEM and Xinchuang (Information Technology Application Innovation) devices means it is no longer just a consumer application that requires users to actively download it, but can become a part of the device itself.
Models can be updated, and Agents can be redeveloped, but a cooperative network that has already entered the device system and connects local data and hardware capabilities requires long-term product adaptation and ecological accumulation.
The existing cooperation between Marvis and manufacturers such as Microsoft and Intel, as well as the ongoing OEM and Xinchuang cooperation, are essentially expanding this network.
04. Conclusion
From this perspective, what Tencent MyApp wants to do is indeed not just an AI assistant.
In the past, MyApp distributed software and apps: distributing software in the PC era and apps in the mobile internet era. In the Agent era, the form of the entry point has changed—users are no longer just looking for an application, but directly proposing an intent.
For Marvis, the next step is to further "complete things on behalf of users."
This is also why the Marvis team values the completion of users' real tasks more.
If this path continues to extend, what Marvis ultimately connects will not be limited to PCs. A PC, a NAS, an AI Box, or even devices like cars and glasses, could all become nodes for Agents to execute tasks.
What Marvis truly wants to reach is the layer between model capabilities and real devices.
But the deeper the device access, the more important permissions, trust, and costs become.
Marvis is currently strengthening this boundary in terms of permissions and key operations:
Cost and privacy are solved by on-device models: through low-bit quantization, Marvis has compressed the 35B-level model from about 30GB to under 10GB, allowing it to run on local devices—data does not leave the local device, and processing does not go through the cloud, which not only saves cloud invocation costs but also keeps privacy in the hands of users.
Permissions and trust, on the other hand, are backed by another set of system mechanisms: when it comes to key operations such as deletion, modification, and payment, Marvis explains first and then confirms; external service access follows the "opt-in by default" principle, and adopts dual confirmation of developer authorization and user authorization.
For an AI that exists in a device for the long term, capability determines how much it can do, but permissions determine whether users dare to let it do it.
This is perhaps where the division of labor differs between an "AI butler" that manages the local environment, understands the user, and keeps to its duties, and other Agents.