从能用到进入生产：腾讯如何构建企业级Agent生产力系统|甲子光年

From Usable to Production: How Tencent Builds an Enterprise-Grade Agent Productivity System | Jiazi Light-Year

From a personal assistant to an enterprise‑level productivity system.

Author｜Tian Siqi

Author | Lizi

At the end of each month, sales managers need to identify which customers have overdue payments and the reasons for the delays. Agentfor enterprises, this is far more complex than checking a single report. It requires connecting the customer management, contract, and financial systems to verify invoices and follow‑up records. For the same issue, sales staff can only see the customers they are responsible for, supervisors can view data of authorized teams, and finance personnel have different permissions.

This is the Tencent Cloud Intelligent Agent Development PlatformADPAn example presented by the person in charge, Ma Yanxiang, at the Tencent Global Digital Ecosystem Conference. Models can already understand instructions and invoke tools, but each enterprise has its own processes and permissions.AgentWhether it can truly complete the work depends on whether it understands these rules and can execute tasks across systems.

AIThe question of whether AI investment can be translated into business value is becoming increasingly prominent.Gartnerin2026OctoberCEOsurvey shows44%of surveyed CEOs find it difficult to demonstrate AI’s financial returns to stakeholders,63%believe that reengineering operating models is a key step to accelerate value realization,52%plan to deploy AI agents as a priority for employees.

10月29日至30日，腾讯全球数字生态大会在深圳举行。腾讯集团高级执行副总裁、云与智慧产业事业群CEO汤道生强调，员工用AI Agent提效已经是必选题，“否则不进则退”。

In his view, the difficulty of enterprise implementationAgentis not that the model is insufficiently powerful, but that there are too many breakpoints:AIcan write proposals and perform analysis, yet cannot connect to business systems or obtain data. The key to implementation lies in connectivity—linking people, tools, systems, and data—so thatAgentunderstands the business and participates in actual work.
To this end, Tencent will focus on three major scenarios—office collaboration, data management, and business execution—to help enterprises build an operable, evolvable intelligent productivity system.

At the same time,At the AI product launch stage of Tencent’s Global Digital Ecosystem Conference, over two days, Tencent released more than 40 Agent products, focusing on the four areas of application, data, security, and engineering.From the daily work entry points used by employees to business application development, cloud operation, and security governance, Tencent is attempting to build a complete product system that supports Agents entering enterprise production.

 

1. Super Individuals Are Not Equivalent to Super Organizations

When preparing for this conference talk, Tang Daosheng handed his ideastoWorkBuddyto organize, then used Yuanbao to verify the information, repeatedly discussing and revising. This way of human‑Agentcollaboration is becoming a part of everyday work.

According to the National Data Bureau, China’s daily averageTokencalls increased from2024beginning of1,000billion, growing to2026yearJulymonth’s over500trillion, increase exceeding5,000times.

However, an effective AI assistant does not automatically grant the organization higher productivity.

WorkBuddySenior product manager Li Yu pointed out that employees’ data are scattered across differentAI tools; switching to another application requires re‑explaining project progress, and companies also have to manage accounts and permissions separately. Many companies have cultivated a group of AI‑savvyAI “super individuals”, yet have not established effective collaboration methods.

“Super organizations aren’t bought, they’re grown.”Li Yu said.

This release’s Buddy One office suite, which is Tencent’s solution. It uses WorkBuddy as a workbench, integrating Tencent Docs, ima and other products, bringing accounts and usage management into a unified system. Tencent Docs allows Agent to understand progress through document content and edit history, ima provides knowledge management, and Tencent Cloud Drive centralizes the storage of files and Agent outputs.

Tang Daosheng described a new collaboration model: after team discussion, the requirement is handed toAI,Agentproduces the first draft, employees revise and accept it, and the result is then distilled into the knowledge base. When the next task starts,Agentcan call existing materials, continue the team’s progress without re‑learning the project background, and employees no longer need to repeatedly pass files.

Beyond office work, management also faces another layer of difficulty.ERP, CRMand financial software have accumulated massive amounts of data, yet each uses its own query method. Tang Daosheng noted that a data engineer may need two or three weeks to produce a business report, managers then ask for details after reviewing it, and updating the report takes another few weeks. By the time the data is ready, the decision‑making window may have already passed.In his view, speed itself is a competitive advantage.

Tencent attempts to useMCP APIs andSkills, to letWorkBuddy to connect existing enterprise management systems. Tencent Cloud’s operation service intelligent agent Panshi XiaoO, enables employees to directly query sales performance, customer status, and payment progress without searching reports across multiple systems; the process is auditable and traceable. Before a sales visit, it combinesCRM data and public information to prepare customer dynamics and communication suggestions, assisting in scheduling meetings; after the visit, it organizes reports, generates to‑dos, and initiates approvals as needed.

Tang Daosheng listsoffice collaboration, data management and business execution as three key scenarios for enterprisesAgent.WorkBuddy has already made it easier for employees to access corporate information, but professional tasks such as customer service, sales, and marketing each have their own business rules. To let Agent takes over these tasks, enterprises need the ability to build and manage agents according to business requirements.

2. BusinessAgent‘s value must be realized in operation

Tencent’s newly released Agent application product, covering multiple stages from agent development to specific industry operations.

Tencent Cloud Intelligent Agent Development PlatformADPhelps enterprises build businessAgent; audio and videoAIplatformWANDintegrates multimodal models to reduce content generation costs; Tencent Cloud Real-time Intelligence supports live conversations and converts interactions into business records; Yuanbao Search provides traceable external information; Tianlai voice large model handles multi-person, multi-language audio; Tencent Qidian Marketing CloudMAGIC AIconnects customer analytics with marketing campaigns, Tencent CloudMallCloud XiaoMo assists merchants in operating private-domain business.

Among them, ADP‘s upgrade reflects Tencent’s thinking on building enterprise Agent.

Ma Yanshang mentioned a client that, as early as ADP 3.0 period built dozens of agents, and this year they want to use Agentto identify supply‑chain risks: retrieve publicly available raw‑material and supplier risk events, combine them with internal data to determine which suppliers and components may be affected, and ultimately generate a deep‑analysis report.

Customers not only demandAgentto operate autonomously, and also care about whether the process is controllable,Tokenwhether the cost is justified.ADPlets the workflow handle fixed batch retrievals, with the cloudAgentperforms cross‑validation and deep analysis, balancing task flexibility with cost.

New versionADPfurther introduces natural language into the build process.Business users only need to describe requirements, and the platform can assist in building agents and workflows; configuration, preview, debugging, and evaluation can also be completed in the same workspace.

The retirement trust scenario presented at the conference is an example: employees raise a need to provide retirement trust consulting for elderly clients, and the platform generates an intelligent agent along with related knowledge, tools, and workflow configurations.ADPIt also integrates the intelligent agent with site services, the former handling analysis and execution, the latter hosting the UI pages, business logic, and data state.Thus, enterprises can directly build complete business applications without being limited to a single intelligent agent.

LetAgentrun, and also ensure it can operate long‑term.ADPthe cloud component of this upgradeAgent Harnessis responsible for managing task execution, memory, tool invocation, and sandbox scheduling.When employees shut down their computers, long‑running tasks can continue executing in the cloud, and sandboxes are started and stopped on demand, reducing idle resources.If business users find the results incorrect, they can directly submit feedback; the platform generates candidate optimization criteria, which are enabled after manual review and sample re‑evaluation.

The built Agent can be distributed by department and position through AgentPortal, and can also enter WorkBuddy. Conversely, employees can validate effective workflows in WorkBuddy, using the ADP connector to transform them into enterprise agents for unified management and promotion. The new ADP opens 130+ Open API, covering ten modules, enabling employees’ accumulated work methods to become capabilities that the team can reuse.

Yili has brought this collaboration to the sales floor.WorkBuddyaggregates business data, formulates strategies, and then distributes them to thousands of sales assistants, boosting overall operational efficiency by over50%;ADPsales assistants’ AI agents provide product recommendations and communication scripts,A/Btesting shows a sales order increase of26%.While faster generation of analytical reports is certainly important, whether the strategies are adopted on the front line determines if they truly generate business value.

Another requirement is to obtain external information promptly and reliably. YuanbaoAIsearch general manager Ye Shani mentioned that enterprise customers oftenfeedbackAIanswers are slow, inaccurate, or even unable to answer the question. The more complex the task the model undertakes, the more important the timeliness and source of the information become.

This upgrade of the Yuanbao search integrates the public web, Tencent ecosystem, and industry sources, using multi-source recall, ranking, and result enhancement to return structured information with verifiable sources.Its index scale reaches the hundred‑billion level, connecting157authoritative vertical domain interfaces, with daily calls in the hundreds of millions; enterprises can useAPI,MCPorSkillto use, without needing to build their own search system.

“Without a source, even a beautiful answer is hard to believe.”Ye Shani said.

Voice enables Agent to handle more business information in meetings and calls. Tencent Cloud Vice President and Tianlai leader Shang Shidong noted that a leading customer is no longer only concerned with recognition accuracy, but also hopes that voice AI helps increase revenue and accurately pinpoint causes when issues arise.

Tianlai released three large models this time: a speech‑understanding model that incorporates meeting documents,screen sharing, participant information and other contextual cues to assist transcription,distinguishing speakers in overlapping and alternating speech,supporting16languages and multiple dialects, with multi‑turn refinement to improve accuracy; the speech‑generation model uses an end‑to‑end architecture,supporting natural conversation and dialect expression,and can clone a voice in 5 to 10 seconds; the simultaneous‑interpretation model can output translations while the speaker continues speaking, preserving the speaker’s timbre. All three models are now available on Tencent Cloud Speech Services.

AgentThe more work an agent takes on, the more enterprises need to address issues of continuous operation, cost control, and team sharing.

3. Installation is just the beginning; only if you can manage it can you scale

On March 6 this year, a long line formed for the OpenClaw installation event downstairs at Tencent Building. Some came with their children, and some retired engineers came specially to try it out. At that time, successfully installing the Agent and connecting to QQ was enough to excite users.

After more than half a year, Xiao Na, Vice President of Tencent Cloud’s General Computing product, found that people were no longer satisfied with just installing the Agent; they wanted to assign a task and receive the result the next day.Tencent’s own development process has also changed: releasing a version in a day used to be difficult, but now some enterprise products can be built and launched by product managers within three hours, with developers focusing mainly on architecture and quality verification.

Tencent Cloud thus integrates and upgrades its cloud-hosted self‑hosted Agent product into AgentHosting, with the brand concept “Always Here”: people can leave their computers, the cloud Agent remains, and tasks keep moving forward.

Around this point,Tencent Cloud proposes Here to Run, Here to Govern, Here to Connect, focusing respectively on continuous operation, organizational governance, and capability connection. According to Tencent’s disclosure, the related products have supported millions of agents running stably in the cloud, serving thousands of enterprises.

This brand covers four products. First, the one responsible for deployment and residency,targeted at individuals and developersLighthouse Agent—itsupports one‑click deployment of agents such as OpenClaw and Hermes on the user’s own cloud instance. After the local computer is shut down, tasks continue to run, and the memory plugin retains context and preferences, reducing the need for repeated explanations.

When faced with complex tasks, users can also involve multiple Agent collaboration: the navigator splits the task and hands it to specialists such as industry analysts, programmers, etc., Agentand then aggregates the results. Each Agentcan be configured with its own model to avoid consuming expensive compute for simple work. After integrating with Tencent Meeting and Tencent Docs, it can also handle everyday office tasks; when encountering steps that require human interaction such as account login or file upload, Agentwill request takeover via a cloud desktop, and resume execution after completion.

But when a company runs dozens or hundreds of agents, the trouble goes far beyond keeping them online. Xiao Na gave a cost example: using a top‑tier model to adjust the font and color of a single PowerPoint slide ended up costing a few hundred yuan, equivalent to hiring a senior design expert for half a day, yet only achieving simple layout. Companies must also define each agent’s responsibilities, permissions, and costs.

This is exactly the unified governance issue that ClawPro aims to solve. The product lead Huang Qiting discovered that some employees want the company to provide pre‑configured agents directly; the skills debugged by experienced staff can only be shared in group chats, and after updates they have to be resent, making it difficult for managers to grasp the team’s overall usage.

Starting public testing in April this year, the officially full‑scale released ClawPro can centrally manage multiple cloud and on‑premise agents, distribute configurations and capabilities by organization, and isolate management models, capabilities, and quotas through a four‑level permission hierarchy. Enterprises can view token consumption centrally, conduct security audits, and also embed agents into business systems via API.

A tax and accounting outsourcing firm organized the experience of senior accountants into Skills, and after distributing them to employees via ClawPro, the number of clients per accountant grew from over 200 to more than 400, boosting overall efficiency by 50%. ClawPro now serves more than 1,000 companies. Enterprises no longer need each employee to set up an Agent from scratch, and mature work methods can be quickly propagated across the organization.

After accumulating experience, continuous circulation is also neededto form capability supply. Under AgentHosting,SkillHubprovides skill discovery, evaluation, updating, and distribution capabilities, enabling developers and enterprises to find and reuse existing professional skills; SkillPay further connects skill payment and delivery, allowing valuable experience not only to be shared but also to become a sustainable supply.

The other end of the product system is professional scenariosDeep cultivation. Tencent Cloud’s open‑source, self‑hosted AI Agent engineOctop, starting from cloud server operations and maintenance, extends Agent capabilities to concrete work. After more than three months of open‑source, its GitHub stars have already exceeded 11,000. In the community, a legal‑tech company has built a contract review workbench based on Octop and, by integrating with the Chang’an Chain, provides notarized evidence for AI citations.

From Lighthouse Agent enabling continuous task execution, to ClawPro’s unified governance and SkillHub’s circulation of professional capabilities, and further to Octop’s deep industry scenarios, AgentHosting aims to solve not just how convenient it is to install an Agent, but how to keep a large number of Agents working long‑term, continuously acquiring new capabilities, and integrating into the enterprise’s existing business systems.

For enterprises, the real value is not how many agents are deployed, but whether they can be turned into a manageable, collaborative, experience‑accumulating digital workforce.

4.AgentEnter production, ultimately you need to settle the operating accounts

AsAgentstarts participating in corporate business decisions; companies must not only let it access data but also ensure it correctly understands the meaning of the data. The same sales figure may be measured with different criteria by different departments.AgentEven if it can query the database, it may still fail to accurately grasp the business.

Tencent this time releasedAInative multi‑model databaseTDSQL Nexa Lakebase, throughNexa Knowledgebuilds a sustainably evolving semantic layer, helpingAgentunderstand enterprise data and its business relationships. Multi‑modelAIdata foundationTDSQL Boundlesswith a single kernel that handles transactions, wide tables, analytics, search and semantic retrieval;TDSQL‑C PostgreSQLprovides a flexible, isolated database backend;TencentDB Agentfurther packages years of Tencent Cloud database operation experience intoSkills, allowingAgentto take on part of theDBAwork.

After correctly interpreting the data, it is also essential to ensure that the usage process is safe and controllable. The seven security sessions at the AI Product Release Platform extend from network offense‑defense to business risk mitigation. Tencent Cloud AI Agent’s end‑to‑end network security solution can unify user and Agent identities, implement fine‑grained authorization, and detect risks in multi‑turn conversations, enabling data masking, human takeover, or blocking on demand. This allows the Agent to obtain the permissions needed to complete tasks while operating within clearly defined security boundaries, helping enterprises confidently entrust more business to AI. The next‑generation Cloud Security Center integrates host, container, and cloud security posture management, shifting protection from reactive response to proactive operation.

TencentCloudUnlimited UtilizationAIconducts independent penetration testing.According to Tencent, the average verification time per site is about91minutes, costing a few hundred yuan, helping enterprises discover vulnerabilities more efficiently.On the business side, the Tianyu Risk Control Agent, combined with black‑gray market intelligence, can perform business risk analysis and policy tuning anytime, anywhere through AI‑driven conversation.

These fundamental capabilities may not directly drive sales growth, but they determine whether a company can confidently entrust more business toAgent. Tencent’s advantage lies in the fact that products such as WeChat Work and Tencent Docs have already become part of daily collaboration in many enterprises, and Tencent Cloud has accumulated experience serving customers across various industries. This gives Tencent the opportunity to start from the work portals familiar to employees, further connect business systems, and participate in the company’s operational processes.

Looking back at Tencent’s 44sessions of centralized releases, from application building to data, security and engineering capabilities, its layout has already exceeded a single AItool category. The operable and evolvable intelligent productivity system proposed by Tang Daosheng, symbolizes that Tencent is also gradually deepening its participation in the daily operations of enterprises.

AgentWhether it can truly become a productivity driver for enterprises ultimately must be answered by business outcomes.

For Tencent, the important thing is not just getting more enterprises to deploy Agent, but to have customers willing to entrust it with more critical work and reap benefits over long‑term use. Whether professional expertise can be continuously accumulated within an organization, and whether complex tasks can be reliably completed by the Agentwill determine whether AItruly goes beyond personal efficiency gains to fundamentally change how enterprises operate and compete.

(Cover image source: Tencent Global Digital Ecology Conference)

END.