Tencent Cloud WorkBuddy Upgrades Four Enterprise-Grade Capabilities; Executives Respond to Agent Security and Cost

Smart Things (Zhidongxi) reported on October 28 that today, at the 2026 Tencent Global Digital Ecosystem Summit, Wang Huixing, Vice President of Tencent Group and CTO of Tencent Cloud, announced four core capability upgrades for WorkBuddy's enterprise-grade services, including open platform capabilities, enterprise-grade security solutions, an observability center, and WorkBuddy Managed Agent (WMA). Tencent Cloud Agent Runtime, the underlying infrastructure supporting these capabilities, has also been upgraded.

After this upgrade, WorkBuddy will support connecting to enterprise systems and third-party model integration. On one hand, it integrates with Tencent office products such as WeCom and Tencent Docs; on the other hand, it connects to third-party collaboration platforms as well as enterprise business systems such as ERP and CRM through MCP and connectors. At the same time, in addition to Tencent Hunyuan, WorkBuddy also supports integration with third-party open-source models and enterprise self-deployed models.

Addressing the security and cost concerns commonly held by enterprises, WorkBuddy provides capabilities such as identity and permission control, sensitive data protection, and operation auditing, and uses the observability center to track Token consumption, task execution traces, and resource usage across different business scenarios.

In addition, WorkBuddy Managed Agent (WMA) has also officially opened to enterprises and individuals. This service moves the agent's task execution capabilities to the cloud, allowing enterprises to create digital assistants based on their own business processes and, with the runtime environment provided by Agent Runtime, enabling agents to continuously execute tasks and schedule computing resources on demand.

Wang Huixing revealed that currently, for Tencent Cloud's training scenarios alone, the daily sandbox creation volume reaches 14 million times. Since WorkBuddy launched in March this year, it has cumulatively completed 9 billion user task requests, and Token consumption has grown 50 times compared with the initial launch.

During the summit, Wang Huixing, Vice President of Tencent Group and CTO of Tencent Cloud; Liu Yi, Vice President of Tencent Cloud and head of Tencent CodeBuddy and WorkBuddy; and Dong Zhiqiang, Vice President of Tencent Cloud and head of Tencent Security, had a nearly two-hour exchange with media including Smart Things, responding to questions on enterprise AI implementation, Agent security, model competition, cost control, and the next generation of agent forms.

In the group interview, the three revealed a number of noteworthy points:

1. The bottleneck for enterprise AI implementation is shifting toward business context. Wang Huixing believes that models and Harness (the agent orchestration framework) have already crossed a threshold at the product and technical levels, and the next focus is to enable Agents to connect to enterprises' real data, permissions, and business processes.

2. Enterprise AI still lacks a recognized core measurement metric. Wang Huixing mentioned that business growth and reduced operating costs are the ultimate outcomes, but the industry has not yet formed a unified North Star metric for measuring improvements in organizational productivity.

3. WorkBuddy is focusing on optimizing two metrics: task completion rate and hidden error rate. Liu Yi revealed that in addition to whether a task is completed, it is also necessary to check for data errors, formatting issues, and consistency contradictions, though specific metric data has not yet been disclosed.

4. Agent privilege escalation and prompt injection are real security risks. Dong Zhiqiang mentioned that WorkBuddy restricts Agent behavior through methods such as tool-call whitelists, sandboxes, operation auditing, manual confirmation, and rollback mechanisms.

5. Proactive Agents will become a key direction for 2027. Liu Yi revealed that WorkBuddy will strengthen proactive perception of work signals such as emails, WeCom messages, and periodic reports.

6. Tencent is advancing an enterprise digital assistant platform. Enterprises can reuse the execution, security, observability, and evaluation capabilities accumulated by WorkBuddy to build professional digital assistants for legal, finance, HR, and other functions.

7. Tencent will not cut complex Agent tasks just because computing costs are high. Liu Yi expects that with falling Token prices and engineering optimization, the relevant business may gradually form more significant gross margin space in the next two to three years, but this judgment remains uncertain.

The following is an organized and summarized version of the key Q&A, edited by Smart Things without changing the original meaning:

I. Models and Harness have crossed a threshold; enterprise context is only the starting point

Q1: At present, where exactly is the biggest obstacle for enterprises in moving from adopting Agents to large-scale application?

Wang Huixing believes that enterprise Agent implementation needs to be viewed at two levels: on one hand, technical and capability issues that products such as WorkBuddy themselves need to solve; on the other hand, issues of internal business management and organizational support within enterprises.

Regarding the product level, he summarized four capabilities that need to be prioritized.

The first is connecting to real business. Agents must not only connect to WeCom, email, and Tencent Meeting, but also be able to connect to enterprise management systems such as Kingdee and Yonyou, as well as third-party collaboration platforms from competitors.

The second is security governance, including Agent behavior control, privilege escalation protection, data leakage prevention, and operation auditing.

The third is effectiveness measurement. Enterprises need to know who uses Tokens, where they are spent, and what business results they produce.

The fourth is large-scale deployment. In addition to providing personal assistants for employees, it is also necessary to support enterprises in creating professional digital assistants for legal, HR, finance, and other functions, and to distill business experience into capabilities that can be repeatedly invoked.

Wang Huixing said: "In essence, models and Harness have already crossed that threshold at the product and technical levels. What we do next is prepare the context."

Q2: Will the data silo problem that was not solved in the previous round of enterprise digitalization be easier to solve in the Agent era?

In the past, different SaaS systems within an enterprise usually had their own data formats, data definitions, and interface specifications. To make two systems work together, a significant amount of development and adaptation work was required. Now, AI can understand the data schemas and interfaces of different systems and generate converter and adapter code based on business requirements.

In Liu Yi's view, this means that Agents have an opportunity to reduce the integration costs between enterprise systems, enabling data and processes originally scattered across different SaaS systems to collaborate more smoothly.

Q3: Where does the long-term competitiveness of enterprise AI office solutions actually come from—models, Harness, or the connection ecosystem?

Liu Yi believes that the connection ecosystem is one of the most important capabilities of an enterprise AI platform. He explained that enterprises have a large number of internal business systems, data, permissions, security, and approval rules. If an Agent cannot fully connect to these systems, it cannot string together a complete business process.

In addition, enterprises also have upstream and downstream relationships within industry chains. Different enterprises possess their own professional experience and business capabilities, and in the future these capabilities can not only serve internal employees but also serve external enterprises through Agents.

Wang Huixing added from the perspective of models. He believes that continued improvements in model capabilities may indeed reduce some of the complexity of Harness and product engineering, but greater model intelligence often means greater computational cost.

Taking PPT creation as an example, not all tasks require calling the most powerful model. Through collaborative optimization of Harness, product capabilities, and models, the expected results can be achieved at more appropriate computational cost.

Therefore, Tencent will, on the one hand, continue to invest in improving the capabilities of the Hunyuan model, and on the other hand, explore joint optimization of models, Harness, and specific business scenarios.

Q4: When third-party partners put their Skills and business data on WorkBuddy, how can they prevent their core capabilities from being obtained by the platform or other users?

Wang Huixing responded that some partners have already raised similar concerns, for example, whether quantitative trading-related Skills, after being connected to WorkBuddy, might be obtained by other users.

He introduced that WorkBuddy is addressing this issue through cloud hosting and encryption mechanisms. Partners can deploy Skills, MCP services, or core data sources in the cloud rather than directly exposing complete capabilities on the client side. At the same time, through methods such as end-to-end encryption, the platform and users are restricted from directly accessing partners' core data and technical accumulation.

For personal or enterprise data, WorkBuddy is also building a unified repository that supports connections to local disks, enterprise cloud drives, and other knowledge sources, and on this basis provides retrieval, analysis, vectorization, and memory capabilities.

Q5: How can one determine whether an enterprise has truly used Agents to improve organizational productivity? Is there already a unified metric?

Wang Huixing believes that ultimately it still depends on business results, such as whether the enterprise has achieved growth or can complete the same work at lower operating costs.

Rather than directly requiring enterprises to provide a clear ROI, he believes that at the current stage a more important step is to first address security concerns so that frontline employees can actually use AI. This is because the specific business bottlenecks of many enterprises may not necessarily be identified when management plans in advance, but may exist in the tasks that frontline employees handle daily.

When employees begin to use Agents to accomplish things that were previously difficult to complete, enterprises can further identify which business links have generated value and gradually establish corresponding measurement standards.

Wang Huixing also mentioned that in the future Agents may change the allocation of human resources and organizational structure within enterprises, but these changes will require some time before they can be observed and evaluated.

Q6: How does the WorkBuddy observability center evaluate the actual effectiveness of Agents? What key metrics are currently available?

Liu Yi revealed that WorkBuddy is internally focusing on optimizing two metrics: task completion rate and hidden error rate.

The task completion rate measures whether the Agent has completed the task assigned by the user and is a relatively direct metric.

Liu Yi emphasized that superficial completion of a task does not mean there are no problems with delivery quality. For example, an Agent may generate a report but introduce errors into the data; it may also complete a PPT but have issues with formatting, layout, or consistency.

These problems may not cause the task to fail immediately, but they affect final delivery quality, so they need to be further measured through the hidden error rate.

Liu Yi said that the team has established multiple benchmarks around these two metrics and uses evaluation results to continuously optimize the delivery capabilities of models and Agents.

As for specific data such as task completion rate and human intervention ratio that media follow-up questions asked about, he said it is not convenient to disclose them for the time being.

Q7: Do risks such as unauthorized Agent access, prompt injection, and even loss of control exist in real enterprise environments?

Dong Zhiqiang stated clearly that such risks do indeed exist.

He explained that the operating logic of traditional software is usually constrained by code written by developers, and its behavior paths are relatively fixed. Agents, however, need to autonomously choose tools and execution steps according to goals, which means they may attempt operations beyond expectations in order to complete tasks.

One such risk comes from prompt injection. For example, when an Agent reads web pages or emails, malicious content may carry hidden instructions that induce it to ignore original requirements, access other data, or perform unauthorized operations.

Therefore, security systems must treat external content such as web pages and emails as data, and must not directly treat the content within them as executable instructions.

Another risk is unauthorized access. Dong Zhiqiang gives an example: although an employee has access permission to a certain CRM system, the Agent may, through different interfaces or call paths, access data beyond what is actually needed for the work.

To address this, a whitelist can be set at the AI gateway layer for the tools and parameters that the Agent can call, and the read and write permissions of interfaces can be restricted according to business needs.

In addition, high-risk operations also require accompanying manual confirmation and rollback mechanisms, rather than relying entirely on the model's own judgment.

Q8: Enterprises have strict data security requirements. How does WorkBuddy control data flow and execution permissions?

Dong Zhiqiang introduced that WorkBuddy's security capabilities cover enterprise accounts, tool and Skills detection, sandbox execution, data protection, and final artifact management.

At the data level, enterprises can perform classification and grading checks on sensitive files, and issue reminders, desensitize, or encrypt sensitive information such as mobile phone numbers and ID card numbers.

For enterprises that require data not to leave a specific environment, WorkBuddy supports solutions such as self-deployed models, VPC dedicated networks, and confidential computing.

At the behavioral level, the commands executed by the Agent, the models called, external tools, and final results can all enter the audit system, facilitating post-incident investigation.

For operations that may modify important data, they can first be executed in a sandbox, and then a human confirms whether to write to the real business environment.

Dong Zhiqiang also mentioned that in the future, personal digital assistants will also need to have independent Agent IDs and clarify permission boundaries through an identity system, rather than infinitely inheriting all permissions held by the user.

III. Betting on proactive Agents in 2027, the difficulty is getting AI to understand long-term goals

Q9: From executing a single task to proactively completing work, how will WorkBuddy evolve next?

Liu Yi believes that an important progress of WorkBuddy in 2026 is the development from executing a single task according to user intent to completing multi-step, long-cycle tasks around user goals.

The biggest difference between the two is that the Agent not only needs to execute operations, but also needs to check whether the results achieve the goal the user truly wants. For example, when making a PPT, in addition to accurate content, aesthetics, layout, and the final presentation effect also need to be considered.

On this basis, Liu Yi revealed that in 2027 WorkBuddy will focus on strengthening Agent proactivity.

In office scenarios, the Agent can perceive new emails received by users, collaboration messages in WeCom, or periodic work needs such as quarterly reports, and proactively prepare reports, remind about to-dos, or provide suggestions at appropriate times.

He believes that this capability is not only applicable to personal office work, but can also enter enterprise business processes and take on more automated work.

Q10: Why have proactive Agents not yet been deployed at scale? What is the biggest technical challenge?

Liu Yi divides the difficulty into two aspects.

First is trustworthy hosting. Proactive Agents need long-term access to user-authorized data and tools, so they must solve issues such as identity and permission protection, credential isolation, security guardrails, and prompt and Skills attacks.

If users do not dare to hand data and permissions to the Agent, continuous hosting cannot be achieved, and true proactive execution is out of the question.

Second is long-term goal understanding and continuous evaluation. Liu Yi believes that the Agent must not only know what task the user has currently assigned, but also be able to understand the long-term goal behind the task, judge whether the goal has been completed, and determine when to continue acting next.

This requires the model and the Harness to jointly build capabilities for long-term goal extraction, state tracking, and result evaluation. Unlike traditional scheduled tasks, proactive Agents need to identify the timing of actions on their own, rather than relying entirely on trigger conditions preset by users.

Q11: When Agents can directly generate interfaces and call software, will traditional office software still be needed in the future?

Liu Yi believes that in the future, software capabilities may further develop toward modularization and plug-inization. Functions originally provided by complete applications can be split into more independent and composable capabilities, which are then selected and called by the model according to user needs.

However, Wang Huixing added limitations regarding privacy and security.

He believes that there are differences between enterprise internal systems and personal life applications. Enterprises can open business interfaces under a clear authorization framework, but when personal chat records and private data are involved, one cannot consider only whether the Agent has the operational capability.

IV. Digital assistants become the focus of the enterprise edition, and SMEs mainly rely on partners for delivery

Q12: Why does WorkBuddy need to further extend from a personal assistant to an enterprise digital assistant?

Wang Huixing confirmed that digital assistants will be an important direction for the next stage of the WorkBuddy enterprise edition.

He explained that personal assistants mainly provide help around employees' daily work, while enterprise digital assistants need to understand the business context, data, and processes of a certain professional field. For example, enterprises can create professional digital assistants for different positions such as legal, finance, HR, or business management.

In this process, enterprises do not need to rebuild the entire system from open-source models or Harness, but can directly reuse the security, cloud operation, observability, and Agent evaluation capabilities accumulated by WorkBuddy.

Wang Huixing also mentioned that Tencent will export to enterprises the evaluation system used by WorkBuddy to help enterprises test the digital assistants they build.

For specific industry applications, Tencent will choose areas it excels in to participate in building, but it does not intend to take on the development of digital assistants across all vertical domains.

On internal practice, Wang Huixing introduced that Tencent has already been exploring digital assistants in its business management systems, allowing employees to query business data, handle approvals, and perform other tasks through natural language.

Q13: Large enterprises need FDE on-site support; how can SMEs obtain similar services?

Liu Yi said Tencent is adopting a differentiated delivery approach.

For large enterprises, business systems and internal processes are often very complex, and Tencent needs to send FDEs (Forward Deployed Engineers) on site to understand the enterprise's roles, SOPs, data, and permission systems, and then integrate Agent capabilities into business processes.

But this model cannot be directly replicated for all SMEs. Liu Yi said: "SMEs rely on partners, not on FDEs; only through partners can it scale."

He revealed that Tencent is building an FDE training system to enable ecosystem partners to master the corresponding Agent delivery capabilities and provide SMEs with remote support, lightweight deployment, and necessary on-site training.

These partners need to undergo relevant Tencent Cloud training and certification, and accept assessment and supervision. If a service partner violates security or compliance requirements, Tencent will take measures such as blacklisting.

Q14: After Tencent internally adopted AI programming at scale, what changes have occurred in the role division and organizational structure of R&D teams?

Liu Yi revealed that WorkBuddy and Tencent Cloud's internal AI-native R&D team began exploring the deep introduction of AI into the R&D process from late 2024 to early 2025.

One end is architecture and requirements design. Engineers need to be able to clearly describe problems and design technical solutions, rather than simply handing vague requirements to the model to guess.

Liu Yi believes that this trend of talent capabilities shifting toward both ends is also driving team organizational structures to become flatter. Team leaders and engineers can communicate directly around architecture, solutions, and user feedback, reducing the layers of reporting and requirement transmission of the past.

He mentioned that Tencent will also use FDEs to share internally accumulated AI R&D practice experience with enterprise customers.

V. Not cutting high-cost Agent tasks; commercialization may take another two to three years

Q15: How will WorkBuddy and Tencent's internal products such as Yuanbao and WeCom divide boundaries in the future?

WorkBuddy mainly focuses on enterprise productivity, providing employees with a personal work assistant and helping enterprises build professional digital assistants and corresponding operations platforms.

Between WeCom and WorkBuddy, the emphasis is more on complementary capabilities.

WeCom provides business context such as organizational communication and employee-customer interaction, and WorkBuddy uses this information to execute tasks and deliver results.

Wang Huixing believes that in the future, Agents for personal life and enterprise productivity may form different product systems, and may not necessarily converge into a single entry point.

Q16: Will WorkBuddy rely excessively on Hunyuan? Will Tencent also conduct joint optimization with third-party model vendors such as DeepSeek?

Wang Huixing said that the CodeBuddy team and the relevant DeepSeek team had previously maintained communication at the product, engineering, and model levels.

In his view, different model vendors will optimize end-to-end results and costs for specific scenarios, and cloud platforms and Agent products can also benefit from these optimizations.

Tencent will continue to maintain open access and conduct exchanges and technical collaboration with different open-source model teams, rather than letting WorkBuddy use only a single model.

Liu Yi also mentioned that the optimization goals of models such as Hunyuan, DeepSeek, and Kimi are not aimed only at a certain Agent platform, but at the task scenarios that users actually use.

When models improve their capabilities on high-frequency tasks such as daily office work, content generation, and code development, WorkBuddy can directly introduce these capabilities into its products.

Q17: Agents consume a large number of Tokens to complete complex tasks. As user scale expands, how will the cost problem be solved?

Wang Huixing believes that the economics of the Agent business cannot be judged statically based only on current inference costs.

On the one hand, GPU hardware continues to upgrade, and there is still room to optimize model inference efficiency; on the other hand, through joint optimization of models, Harness, and context engineering, the resource consumption required to complete the same task can also be reduced.

He also mentioned that deep users often have a strong willingness to pay, and the value of enterprise scenarios cannot be measured only by Token cost; it is also necessary to consider whether Agents can complete tasks that previously required more manpower or even could not be completed.

As for whether complex Agent tasks will be limited due to excessively high costs, Liu Yi said: "We have never thought about cutting off some highly complex, high-value scenarios; rather, it is about how to serve them well."

Liu Yi believes that as Token costs decline and technology optimizes, complex problems can be handled at a lower marginal cost in the future.

He expects that when marginal revenue exceeds marginal cost and business scale further expands, Agent services will have the opportunity to form a healthier gross margin structure.

When media asked when this state might be reached, Liu Yi's estimate was: "Maybe two to three years." However, he clarified that this is still a judgment based on trends in technology and cost changes, not a profit result that has already been achieved.

Q18: Competition in enterprise AI office is becoming increasingly fierce. How will WorkBuddy face competition from rivals' mature office platforms?

Liu Yi believes that the enterprise Agent market is still at a very early stage.

He judges that about 90% of enterprises have not yet applied Agents to productivity scenarios on a large scale and fully.

Traditional office platforms already have users and collaboration systems, but for AI Agents to truly connect to enterprise business, they still need to solve issues such as model capabilities, Harness engineering, context connectivity, and implementation delivery.

Liu Yi believes that in the future enterprise-level Agent market, it is unlikely that one platform will cover all customers; different vendors may form their own service scopes based on existing customer bases, product systems, and industry capabilities.

In his view, the competitive outcome at this stage is far from determined.

From this group interview, it can be seen that enterprise Agents are encountering problems completely different from those of personal AI tools.

In an enterprise environment, completing a task is only the starting point: who authorizes it to access data, who checks the execution results, who is responsible when errors occur, and how much benefit these investments actually generate all require a clear management mechanism. This also explains why, even as model capabilities continue to improve, the large-scale implementation of enterprise AI may not accelerate in sync.

These issues mean that there is still considerable engineering and management work to be done before Agents can become enterprise productivity that can independently bear responsibility.