Tencent WorkBuddy recently announced the opening of its cooperative ecosystem, integrating software partners such as Tongda Xin, Guangfa Securities, and Beida法宝, along with hardware partners including Plaud, Rokid, Insta360, and iFlytek, to build an Agent OS work system.

Author and source: GeekPark

Earlier, Anthropic released the Model Hardware Standard, aiming to establish a universal language for AI agents to communicate with hardware. After integrating this standard, devices from different manufacturers with varying interfaces—such as microscopes and robotic arms—can be discovered, understood, and controlled by agents.

This signals a shift occurring in the AI industry. Over the past year, competition centered primarily on models’ reasoning, generation, and multimodal capabilities; now, agents are stepping out of chat windows and into laboratories, factories, and enterprise environments. The industry’s challenges have consequently shifted downward: once models are sufficiently intelligent, who will connect tools, hardware, data, and tasks?

On September 2, Tencent WorkBuddy announced expanded partnerships with more collaborators. Applications and software partners such as Tongda Xin, Guangfa Securities, Peking University Legal Info, Tencent SSV Teacher Assistant, Weimeng, Beisen, Yunzhangfang, and Fanruan have joined the ecosystem. Hardware partners including Plaud, Rokid, Insta360, Youlanzi, iFlytek, Anker, Mammoth, and Jingzao have also contributed to WorkBuddy’s ecosystem.

Why must an AI product connect simultaneously to professional software, industry services, and multiple hardware devices? This list outlines the complete work environment an Agent will face: professional software provides industry knowledge and business actions, intelligent hardware delivers real-time information through sound, images, and mobility, and the Agent integrates them into a continuously progressing task chain.

Over the past year, AI office products have been able to write, search, summarize, and handle increasingly complex content. However, real work remains scattered across different software, devices, and account systems. After receiving AI suggestions, users often still need to manually transfer information, switch tools, and connect workflows.

So, when AI begins to engage in complex tasks, the real competition lies in who can build a workflow system for the Agent—enabling it to remember context, understand task status, coordinate applications and devices, and transform user intent into results within controlled permission boundaries.

In this collaboration, WorkBuddy aims to provide the answer: Agent OS.

After learning to answer questions, AI must also learn to get things done.

Take, for example, an investment manager preparing for an industry research project. He needs to review previous research notes and meeting minutes, retrieve market data, company announcements, and industry information; after the meeting, he must organize key insights, list unresolved questions to verify, assign follow-up tasks, and ultimately consolidate everything into a comprehensive research report.

Each step involves separate tools: meeting content may remain on recording devices, market information comes from professional financial services, documents are scattered across local files and enterprise knowledge bases, and collaboration occurs through documents, emails, and instant messaging apps. AI can help generate a summary, but it may not know what the user previously researched, which issues were addressed in this discussion, or which conclusions still require verification.

This is precisely the main obstacle agents encounter when transitioning from “answering questions” to “getting work done.” Lin Zuolu, Head of the Open Ecosystem at Tencent WorkBuddy, also mentioned at the launch event: “We’ve already crossed the threshold of agents being able to ‘get things done.’ The real challenge is ensuring they never break down—regardless of the device, system, or scenario.”

Agents must handle complex tasks and require continuous context. They must remember what the user has already done, understand the current stage of the task, determine which tools to invoke, and know which data can be accessed and which actions require confirmation. Memory, task state, tool connectivity, identity permissions, data, and outputs together form the essential environment for an agent’s operation.

Early PCs also went through a similar phase. Before operating systems and standardized interfaces matured, software struggled to exchange information seamlessly, and users themselves acted as connectors between different tools. Today, the fragmentation between applications, devices, and agents brings this friction back in new forms.

The so-called Agent OS addresses this layer of issues. While traditional operating systems manage files, windows, and applications, an Agent OS must also manage tasks, memory, tools, and permissions. The model provides understanding and reasoning capabilities, but this system determines how much of the workflow the model can see and how far it can advance tasks.

Within six months of its launch, WorkBuddy completed over 50 version iterations and was implemented across more than 50 industries, including government, education, retail, finance, media, and transportation. According to Tencent’s Q1 2026 financial report, measured by daily active accounts, WorkBuddy has become China’s most popular efficiency AI agent service. Data from third-party analyst firm Analysys shows that, in June, monthly PC visits exceeded 20 million, maintaining its leading market position.

While some AI office products are still validating single-use scenarios, WorkBuddy has already accumulated a sufficiently diverse range of real-world tasks—the foundation for its attempt to build an OS, since an OS is never designed for a single application.

How to connect a workflow from the Buddy app to smart hardware

WorkBuddy’s open ecosystem first focuses on equipping agents with their “hands” and “eyes.”

Applications and software partners such as Tongda Xin, Guangfa Securities, Beida法宝, Weimeng, Beisen, Yunzhangfang, and Fanruan connect to specialized scenarios including finance, law, business operations, human resources, taxation and accounting, and data analysis. The value they bring extends beyond application entry points—they also provide the industry knowledge, business data, and executable tools required for Agents to complete professional tasks.

Taking financial research and analysis as an example, Tongdaxin has encapsulated its 30 years of professional research capabilities into WorkBuddy. Users can ask a single question and receive structured, traceable, research-report-level content within minutes. This application includes 26 expert agents and 14 skills, which autonomously orchestrate Tongdaxin’s real-time market data, financial statements, and sector information via the MCP protocol during operation. Standardized data cards automatically appear during these calls. Furthermore, when data gaps exist, the system explicitly labels them as “Pending,” rather than fabricating or filling in the missing information.

This is precisely the hard constraint AI faces in professional settings. General-purpose large models can generate analyses that appear complete, but in fields like investment research, law, and medicine, "appearing correct" is entirely different from being traceable and verifiable.

The Buddy app can be viewed as a vertical workflow module loaded onto the WorkBuddy platform. Industry partners integrate their expertise, tools, processes, and brand identities into it, providing users with workspaces tailored to their professional needs. Investment managers, teachers, lawyers, content creators, and merchants may each have their own unique Buddy.

Image source: WorkBuddy

The addition of hardware partners further extends the workflow beyond the screen. Devices from Plaud, Rokid, Insta360, iFlytek, and YouLan bring voice, image, spatial interaction, and mobile office data to the Agent; partners such as Anker, Mammoth, and Jingzao complete the terminal entry points for office and creative environments.

The logic of hardware integration is essentially writing a "driver" for Agent OS—hardware handles perception, activation, display, and interaction (equivalent to I/O devices), while WorkBuddy handles understanding, planning, and scheduling (equivalent to the kernel). Manufacturers do not need to build their own AI or alter their product forms; a single integration enables seamless connectivity across hardware, apps, PCs, and the web, allowing accounts, tasks, memories, and outputs to flow effortlessly between devices.

For example, when a user enters a meeting room with a recording device or smart glasses, spoken remarks become task inputs; after the meeting, the content is transcribed and archived, action items are extracted, relevant materials are retrieved, and follow-up emails and collaboration tasks are pushed to the corresponding systems. When the user switches from hardware to a mobile phone or PC, tasks, memories, and outputs continue seamlessly along the same chain.

The Youlanzi co-branded keyboard mic offers an alternative interaction format, integrating the microphone with a numeric keypad—simply press a key to activate and speak to assign tasks to WorkBuddy, without needing to switch to the software window.

Image source: Youlanzi WeChat Official Account

The classic use case for this product is Vibe Coding—developers describe their requirements verbally, and the agent automatically generates code snippets without requiring them to take their hands off the code editor. For content creators and everyday office users, it also serves as an “always-on” voice interface: speak your inspiration instantly, assign impromptu tasks casually, and turn fragmented ideas into structured to-dos or drafts automatically.

For hardware manufacturers, this offers a more practical path to AI. Building a model from scratch—covering model selection, memory management, task planning, tool invocation, and permission control—requires substantial engineering effort. If hardware handles perception, activation, display, and interaction, while the Agent platform manages understanding, planning, and scheduling, manufacturers can focus their resources on what they do best: delivering exceptional product experiences.

WorkBuddy aims to take on this intermediary role. Users see the Buddy app, experts, Skills, connectors, and hardware, while beneath the surface, a scheduling system organizes context and capabilities around tasks.

The more partners that join, the more complete the platform’s understanding of work contexts becomes—this is both the value of the ecosystem and the key variable determining whether Agent OS can succeed.

How does a working system evolve from a product into an ecosystem?

The value of any operating system stems not only from its own features. It must provide developers with stable interfaces, create space for partners to sustain their businesses, and ensure users enjoy consistent experiences across different apps and devices. The Agent OS follows the same logic.

Finance, law, taxation, education, and content creation each have their own knowledge, data, and workflows, with hardware performing distinct sensing and interaction tasks. It is difficult for a single company to cover such a complex range of work environments. The significance of WorkBuddy’s current opening lies in entrusting the platform with shared capabilities such as memory, task management, tool invocation, identity permissions, and artifact management, while allowing partners to contribute industry-specific knowledge, specialized software, and terminal access points.

This also explains why WorkBuddy is evolving from a single product into an ecosystem. An individual agent’s context is typically limited to one application; the more industries and tools it integrates with, the greater its opportunity to understand how work actually happens.

The technical architecture of the open platform further confirms this. In addition to hardware and the Buddy app, the platform simultaneously opens three key modules: Skills, Experts, and Connectors. Developers can encode a set of industry methods or workflows as Skills, encapsulate domain knowledge and decision-making logic as Experts, and integrate their own APIs or tools via Connectors.

This integration is bidirectional: external AI capabilities enter the Buddy marketplace to serve users, while hardware partners can also gain authorization through the platform to bring WorkBuddy’s ecosystem capabilities into their own devices. The platform’s payment capabilities and distribution channels are being rapidly developed—history has repeatedly shown that the design of the economic system matters more than the technical architecture in determining the long-term trajectory of an ecosystem. Whether developers can earn money and partners can receive sustained returns directly determines whether this open system becomes a living ecosystem or merely a one-time collaboration showcase.

Taking a broader view, Tencent’s latest quarterly earnings report summarizes its AI strategy across three levels: intelligence, applications, and infrastructure. HunYuan provides the foundational model capabilities, while products like WorkBuddy and CodeBuddy handle real-world tasks, supported by computing power and cloud infrastructure for model training and application inference. Tencent’s management also describes WorkBuddy as a flexible agent workspace capable of orchestrating different models and skills based on task requirements.

This positioning means that WorkBuddy’s role within Tencent’s ecosystem goes beyond simply packaging HunYuan as an office product. It acts more like a connecting layer between models, tools, and user tasks: absorbing the capabilities of HunYuan and other models at the bottom, linking to specific productivity scenarios at the top, and validating the usability of models and agent engineering through real-world tasks.

The recently released Hy4 Preview makes this workflow more concrete. Tencent positions it for productivity tasks such as coding, office work, and scientific research, and has integrated it with WorkBuddy, CodeBuddy, Yuanbao, and ima. The Hy4 Preview is co-designed with products like WorkBuddy, focusing on optimizing cross-file collaboration, data analysis, and the complete workflow from information processing to delivering documents, spreadsheets, and presentations.

In the past, large model development and product deployment often followed two relatively independent paths: the model team improved capabilities through benchmark tests, while the product team later identified suitable use cases. The relationship between WorkBuddy and Hunyuan, however, follows a different approach: the model enters real-world productivity tasks first, and product-side evaluations, tool invocation results, and scenario feedback are fed back into the R&D process to drive continuous iteration of the model and Agent engineering.

This forms the data flywheel Tencent aims to build. As model capabilities improve, WorkBuddy can handle more complex and longer-chain tasks; an increase in task types exposes real issues in planning, tool usage, and result delivery; continued integration of applications, hardware, Skills, and experts further increases the platform’s contextual density. Product, model, and ecosystem evolve in tandem, mutually reinforcing one another.

The data flywheel also requires clear boundaries. For enterprise users, how task memory is stored, how business data is isolated, and how critical operations are authorized and audited will determine how deeply the platform can integrate into workflows. As WorkBuddy gains more context, it must also assume greater governance responsibilities. Without this foundation of trust, the larger the ecosystem, the greater the potential risks.

From this perspective, an open ecosystem is itself a fundamental condition for the existence of Agent OS. Within Tencent, resources such as HunYuan, Meetings, Docs, Email, WeCom, and cloud services are already in place, while external partners bring specialized expertise, industry software, and hardware access points. If these capabilities can be unified and orchestrated around tasks, WorkBuddy has the potential to evolve from an internal AI productivity tool at Tencent into a broader work platform serving multiple industries and developers.

OS cannot be achieved through a single release. Can an open ecosystem enable agents to accumulate sufficient context density within each user’s workflow, providing models with real signals for optimization and giving partners ongoing incentives to stay—this is a more fundamental issue than technical architecture.

WorkBuddy can lead the way in defining the direction of an Agent OS, but whether this system truly succeeds ultimately depends on its ability to transform Tencent’s models, products, and connectivity capabilities into a sustainable ecological order.