Author | Zhang Xianing Editor | Hu Min
Today, as large model technology shifts from hype to pragmatism, enterprises are no longer satisfied with merely making conversational AI smarter. "Pragmatic" AI agents capable of executing various complex tasks, autonomously planning skills, and integrating into existing business workflows are becoming an essential requirement for enterprises.
During the recently held 2026 Tencent Cloud Partner Conference, the Tencent Cloud Agent Development Platform (ADP) announced a series of significant upgrades, focusing on deepening platform capabilities, enriching the content ecosystem, and strengthening upper-layer applications.
Wu Yongjian, Vice President of Tencent Cloud and Head of Tencent Cloud Intelligence Product and R&D, accepted an exclusive interview with Leiphone. He detailed the evolution path of Tencent Cloud ADP over the past year and the functional highlights of the new round of product upgrades. He also shared his thoughts on how to improve ADP's industry penetration rate through product upgrades and ecosystem layout in the future. Below are his core viewpoints:
I. From "Single Retrieval" to Complex Task Orchestration and Multi-Agent Collaboration
The core of this ADP upgrade lies in the deepening of its technical core.
Wu Yongjian highlighted the upcoming GraphRAG feature. "Unlike traditional RAG (Retrieval-Augmented Generation), GraphRAG organizes knowledge relationships through graph structures. During the retrieval phase, it first completes knowledge association and aggregation, and then passes the context along with the query to the large model for reasoning and generation. This approach is more suitable for complex problems involving cross-document, multi-hop reasoning, and strong associative dependencies, significantly improving comprehension depth and result stability," Wu said.
Another example is the "node rollback" feature launched for workflows: "Our workflow parameter extraction node supports intelligent rollback at any point. When users communicate with a chatbot, if they find that a previous question was answered incorrectly or information needs to be updated, the workflow can automatically roll back to the previous node for correction, greatly improving efficiency. It also supports asynchronous workflow execution for up to 24 hours (Note: After the upstream triggers the workflow, there is no need to wait for execution to complete; it can return immediately and continue processing other businesses), seamlessly embedding into customers' existing business processes. For instance, long-video content analysis has already been implemented among multiple media industry clients."
Multi-Agent is also one of the key capabilities upgraded in ADP. It has evolved into a mature, externally serviceable product capability, supporting both flexible collaboration through free handover and the execution of stable complex tasks via workflow orchestration of Agents.
II. "Dedicated Cloud" Model: Bridging IaaS and PaaS Resources
In addition to technical core upgrades, ADP has introduced a "Dedicated Cloud" deployment model. For enterprises that value data isolation while wishing to retain the elasticity of cloud resources, this model provides a new option between SaaS and pure privatization.
To promote the efficient implementation of this model, ADP has strengthened its linkage with Tencent Cloud's underlying IaaS resources. "In the past, ADP users utilized a shared rental model," Wu Yongjian stated. "Now we have decoupled ADP from the underlying IaaS resources, allowing ADP to be deployed within users' existing underlying cloud resources." The advantage of this model is that it can quickly attract Tencent Cloud's existing IaaS customers to use ADP.
"ADP can not only be seamlessly integrated into Tencent Cloud IaaS (such as CVM, TKE, etc.), but also supports deployment in cross-cloud and hybrid cloud environments. Meanwhile, it supports flexible access to various cloud service middleware, such as the cloud database TDSQL and cloud cache Redis," Wu said. Through these capabilities, ADP truly achieves an "out-of-the-box" agent development experience and rapidly expands market coverage.
III. C-B Synergy: Feeding B-End Commercialization with C-End Validation
Discussing this year's strategic priorities, Wu Yongjian repeatedly emphasized "C-B synergy," integrating the capabilities of the B-end ADP platform with C-end platforms such as QQ Browser and IMA.
C-B synergy has always been a core competitiveness of Tencent Cloud. For example, QQ Browser's engineering-based document parsing and sandbox capabilities have been utilized in multiple capability modules of the enterprise-level agent development platform, such as RAG and Multi-Agent. Conversely, key capabilities such as stability and controllability, accumulated from long-term B-end enterprise services, are reversely applied to C-end products like the browser and IMA, forming a mutually reinforcing positive cycle.
IV. Commercialization and Market Outlook: Ecosystem Competition under High-Growth Targets
Wu Yongjian predicts that industry competition will intensify this year, with price wars beginning to emerge. ADP's response strategy is to upgrade its business model—maintaining private deployment growth while placing greater emphasis on sustainable subscription revenue. "We are shifting from the original pure token-based billing model to a subscription model that sells SaaS, Dedicated Cloud, or private cloud packages." At the same time, leveraging Tencent Cloud's resource advantages to engage in differentiated competition.
He is full of expectations for this year's product commercialization: "We see market demand accelerating its release, which will drive rapid revenue growth."
Below is the dialogue between Leiphone and Wu Yongjian, Vice President of Tencent Cloud, edited and organized by Leiphone without altering the original meaning.
01 ADP Core Evolution: Comprehensive Upgrades in RAG, Workflows, and Multi-Agents
Q: Over the past year, which upgrades to ADP do you consider the most core and critical?
Wu Yongjian: There are numerous upgrade points. According to our statistics, ADP released six major versions and thousands of functional requirements over the past year.
The most core upgrade is undoubtedly the RAG engine. Because ADP has long been deeply engaged in complex enterprise knowledge scenarios, it has obvious advantages in scenarios such as complex large-table Q&A, mixed text-and-image documents, and comprehensive retrieval of multi-source knowledge, with overall accuracy maintaining a leading position domestically. Before the Spring Festival, we will also launch Graph RAG. Unlike traditional RAG, GraphRAG will first model knowledge points and their relationships, and build a more complete context through multi-hop retrieval before handing it over to the large model for comprehension and reasoning. This effectively reduces "hallucinations" in complex Q&A scenarios, making answers more accurate and traceable.
There are also upgrades in interactive capabilities. For example, with the recently launched AI Widget feature, we were among the first to productize rich interactive Widget capabilities in enterprise scenarios, positioning it as a key direction distinct from pure text chatbots. Compared to traditional chatbots that rely solely on text communication, Widgets provide a "rich interactive" approach, which is particularly suitable for enterprise scenarios. For instance, when users inquire about hotel check-in, traditional chatbots often need to repeatedly ask for information through multiple rounds of dialogue. With Widgets, the required information can be consolidated into information cards, allowing users to directly select or fill in the information during the communication process, significantly reducing communication costs and improving the overall experience.
Q: Besides RAG, what are the new highlights in our workflow and Multi-Agent modes?
Wu Yongjian: The improvement in workflow capabilities has also been very rapid. We only started systematically building this capability last year, but its overall performance is currently in the first tier in the industry and has been recognized by customers.
Take the "node rollback" feature as an example. In a workflow, there are usually several parameter nodes. For instance, in the hotel intelligent customer service mentioned above, these include name, gender, ID number, etc. Traditional workflow orchestration requires inputting in node sequence; if any step goes wrong, you have to start over. However, our workflow supports input in any order. When users communicate with the chatbot and find that a previous answer was incorrect or information needs updating, they can immediately "roll back" to the previous node for correction, greatly improving efficiency. At the same time, we support asynchronous workflow execution for up to 24 hours, which can be seamlessly embedded into customers' existing business processes. In time-consuming scenarios like long-video content analysis, this feature has been implemented in actual business and adopted by multiple media industry clients.
Multi-Agent capabilities are also rapidly maturing on the ADP platform, providing external services through a stable productized form. Users can flexibly define collaboration and handover rules between different Agents through handover descriptions or workflow orchestration, realizing diverse scenario needs from simple to complex. Based on Multi-Agent, ADP also supports creating applications via natural language. After generation, users can directly download standard application packages for secondary configuration and continuous optimization.
Additionally, Tencent Cloud ADP recently launched AI-native Widgets, making the creation of interactive components simpler. Whether through templates, code, or directly describing requirements in natural language, usable Widgets can be quickly generated, significantly lowering the development threshold. Even without a front-end development background, one can complete component creation with a single sentence, bringing ideas to life faster. Combined with Multi-Agent and workflow capabilities, developers can freely combine different agents and Widgets like building blocks, constructing agent applications with more user-friendly experiences and richer forms.
02 C-B Synergy: Feeding B-End Commercialization with C-End Validation
Q: You have mentioned multiple times the need to "enrich content" and collaborate with ecosystem partners and creators to develop high-value Agents. What are the specific plans?
Wu Yongjian: Let's break it down into several aspects. First, through modules like the Model Square and Plugin Square, along with the AI Widget capabilities mentioned earlier, we help users develop Agents efficiently. Currently, the official ADP plugins have expanded to nearly 160, covering general business scenarios and multiple industry verticals. Recently, we have also newly integrated a batch of Tencent-featured plugins, including WeChat Pay, News Fact-Check, Youtu-VITA, and Hunyuan 3D generation.
Additionally, we hope to enable more ecosystem partners to use ADP to build high-quality Agents and industry solutions. Selected Agent applications and solutions can be placed in our cloud application market for external sales. For example, in the media industry, Tencent Cloud ADP has collaborated with ecosystem partners to implement multiple industry applications. During the National Games this year, Guangdong Television produced over a hundred hit content pieces through a media content processing Agent, improving overall efficiency by 40%. Currently, our ADP platform has been implemented in over 20 industries, including finance, media, retail, and healthcare, and the number of ADP partners has grown more than threefold in a year.
Q: Could you elaborate specifically on C-B synergy?
Wu Yongjian: A unique aspect of Tencent's AI product development is that we do not just "assume scenarios" on the B-end. Instead, we continuously refine underlying capabilities in real, complex environments through C-end products, and then systematically productize and output these capabilities to the B-end. Because judging product effectiveness is much simpler through the C-end, such as looking at activity and interaction counts; relying purely on the To-B ADP to refine and validate products would be much slower.
Developers can develop and refine high-quality Agents on the ADP platform. Meanwhile, we support users in quickly experiencing the latest and most promising Agents through C-end experience channels. Agents validated by the C-end market and receiving user interaction feedback will further flow back to the ADP platform for enterprise customers to customize, purchase, and apply at scale.
C-B synergy has always been a core competitiveness of Tencent Cloud. Previously, we applied IMA's algorithm and engineering-integrated document parsing, as well as QQ Browser's sandbox capabilities, to multiple capability modules of the enterprise-level agent development platform, such as RAG and Multi-Agent. Conversely, key capabilities such as practical technology, product stability, and controllability accumulated from long-term B-end enterprise services are reversely applied to C-end products like the browser.
03 Developing High-Value Agents to Build an "AI Enterprise Portal"
Q: Have you or your customers developed any Agents through ADP that truly achieve "making the application work with a single sentence"? Is this proportion high?
Wu Yongjian: Agents with such strong execution capabilities are still in the early stages of development. Currently, the most widely applied mode in the market is still "conversational Q&A," and the capability ADP has accumulated the most in is still RAG. However, the direction of making Agents "do the work" is definitely correct, and similar applications have already emerged, such as in scenarios like hotel customer service.
Q: What characteristics should a high-value AI Agent possess in your mind?
Wu Yongjian: This is hard to define. For enterprise-level Agents, those that can seamlessly embed into business workflows and solve real enterprise pain points are all high-value Agents.
Previously, we collaborated with a leading domestic service enterprise. Through Tencent Cloud ADP, we upgraded and created "Xiao Xi," an "Online Domestic Service Butler" available 7x24 hours. It not only possesses natural conversational abilities but can also directly participate in business processing. When a user makes a request like "find a nanny," the "butler" can respond within 5 seconds, automatically recommend suitable domestic services based on the user's historical information, and flexibly handle pre-sales consultations, service process follow-ups, and after-sales issues, with an overall Q&A accuracy rate exceeding 80%.
In the on-site service phase, we further combined ADP's multimodal capabilities to build a service quality inspection Agent. It conducts automated inspections of service personnel's portrait photos, equipment photos, uniform photos, and highlight photos of the service process, achieving an inspection success rate of over 90%. This effectively reduces manual review costs while ensuring service quality, achieving significant cost reduction and efficiency improvement.
Q: If an enterprise wants to build an independent "working" Agent like Manus on its own, can it be achieved through ADP?
Wu Yongjian: Enterprises indeed hope to integrate some good internal tools or Agents, making calls based on an existing unified entry point, somewhat similar to an AI enterprise portal. We are also making similar product plans, but it will take time to continuously refine and validate. I think there is an opportunity.
04 Building "Ecosystem-Level" Leadership: Upgrading from Product Competition to Ecosystem Competition
Q: From the perspective of current market competition, where does ADP stand? What are the competitive advantages and commercialization strategies?
Wu Yongjian: In terms of product capabilities, in key areas like RAG and workflows, the overall performance is in the first tier in the industry, while in terms of market share, we are in a stage of rapid catch-up.
We see that the commercialization strategies of various companies are not entirely the same. Some focus mainly on pure private deployment projects; their revenue might grow very large at once, but users may not actually use it effectively. There are also some companies that target mid-market or more downstream customers. Through domestic and international channel ecosystem construction, they penetrate the market rapidly like "capillaries," and their revenue is also very considerable. We are exploring in different directions. While growing overall revenue, we also hope to increase the proportion of subscription revenue. This is a point we value highly in our business model.
Another key is to build up the content, which is the true commercial barrier. With this foundation, we hope to introduce more partners. Based on their know-how, we can broaden and deepen our ADP, such as creating exclusive Agents and industry benchmark solutions for various industries. My personal feeling is that model leadership might only be a lead of ten days or a few weeks. Product leadership is probably a lead of one to three months. However, ecosystem leadership can often form a cross-cycle advantageous lead.
Q: We noticed that during the platform rollout of ADP this year, B-to-B synergy was specifically mentioned. Could you elaborate on this?
Wu Yongjian: B-to-B synergy refers to the collaboration between ADP and underlying cloud resources, including the upcoming integration with Tencent Cloud's CVM, TKE, and products like Lighthouse.
Because in the public cloud form of ADP, all resources are already encapsulated, and customers adopt a rental model where they can make calls after purchasing a package. However, today, many customers, especially top-tier enterprises and government clients, prefer to adopt the Dedicated Cloud model, hoping to deploy ADP on their own cloud resources.
Therefore, we decoupled ADP from the underlying IaaS, allowing customers to reuse their existing IaaS to connect with ADP without repeatedly purchasing underlying resources. Partners are also very active because many of their existing customers have already purchased Tencent Cloud products like CVM and TKE. Through integrated sales and delivery, it is easier to drive ADP sales.
Q: Will many customers prefer the Dedicated Cloud solution?
Wu Yongjian: I think there will be. Many large customers lean towards Dedicated Cloud, and many large customers have already purchased IaaS resources on the cloud. They hope to build software on top of it, improving overall ROI through resource reuse.
05 Building a Healthy and Sustainable Business Model with a Subscription System
Q: Does Tencent Cloud ADP have any specific KPIs this year?
Wu Yongjian: The core still focuses on commercialization results. For To-B business, revenue is an important measurement standard, including the scale of enterprise customers and the proportion of subscription revenue, which are all key indicators we focus on.
But revenue is more of a phased result; the true core lies in continuously improving product power, especially in "enriching" the platform content. Whether in public cloud or private deployment scenarios, only by building a sufficiently solid and reusable content ecosystem can a positive cycle be formed, continuously driving commercial growth and achieving long-term, sustainable development.
Q: What considerations are there regarding the billing model?
Wu Yongjian: That's a good question. In the past, Agents were billed by tokens, and the pricing was set too low, which is not a sustainable model. Therefore, we have recently adjusted our charging method, with the core being to abandon pure token billing and switch to a subscription billing model.
Let's break it down into two parts: one is the SaaS end, where enterprises pay an annual basic package fee. If they need to add resources like DeepSeek or buy other content, they can purchase resource points separately; the other is privatization, upgraded to the Dedicated Cloud model, not the original pure privatization. Enterprises purchase our software licenses on TKE, calculated by nodes, and subsequently renew subscriptions normally every year. API services and content on the platform are procured additionally.
Simply put, whether it is public cloud SaaS or Dedicated Cloud now, it is a package-based charging and annual renewal subscription model, not the polarized model of the past where it was either purely token-based or pure privatization. Essentially, it is a shift from a "computing power consumption-based business model" to a "platform subscription-based business model."
Q: What characteristics will market competition present this year?
Wu Yongjian: Market competition this year will further accelerate. On the one hand, more and more enterprises are increasing their investments in algorithms, model capabilities, and computing resources, driving the continuous evolution of Agent technological capabilities; on the other hand, as the number of participants increases, the industry will gradually enter a stage of refined competition.
In this context, relying solely on low prices or single-point capabilities is no longer sufficient to form long-term advantages. Our focus is shifting from "basic capability competition" to "competition in comprehensive product power and service value." By building more comprehensive platform capabilities, more mature industry solutions, and more sustainable service models, we aim to construct long-term competitiveness.
Q: From the perspective of the market prospects for Agents, do you favor the C-end or the B-end? We see that C-end token consumption is huge today, but it is hard to establish a business model, whereas revenue is seen more quickly in enterprise scenarios.
Wu Yongjian: Actually, both the C-end and B-end have good market scenarios. There are many C-end users who are willing to pay for products that generate user value. To-B also has very good market prospects. Things like "AI enterprise portals" and industry vertical Agents will gradually land, and industry solutions will become a new track. In China, it depends on who can truly run through the business model first.
Q: In the field of agents, what other new technologies or trends do you think are worth paying attention to?
Wu Yongjian: From a macro trend perspective, the focus of AI agents is shifting from "being able to chat" to "being able to do the work." Last year, people might have been debating whether Agents could write code or book flights (To-C scenarios). In the next two years, the main theme of the market will definitely be "ROI and reliability." Enterprises no longer just look at demos; they look at whether Agents can truly enter production environments and handle non-standardized exceptional situations. Moreover, rather than just being a program that maintains a leading overall accuracy rate domestically, it should be able to summarize "SOPs" based on failure experiences, achieving self-evolution and becoming smarter with use.
Tencent Cloud's layout is also very clear: we do not just want to build a tool, but to build a "stable productivity" platform. In particular, we need to make good use of the connectivity of our products to let Agents truly enter enterprise business workflows, making the ROI of AI visible and tangible.