Author | Wang Han, Editor | Mo Ying
Today, September 16, the Intelligent Economy Forum was packed to capacity. Experts and scholars from both industry and academia gathered to discuss one single topic: What exactly has AI (Artificial Intelligence) agents brought to the industry? Entering 2026, domestic and international AI giants have turned AI office applications into the hottest B2B battlefield. Products such as personal exclusive workstations, digital human colleagues, and enterprise agents are emerging one after another, frantically seizing every link in the office workflow. Amidst this frenzy, Baidu AI Cloud has put forward a new judgment: The office track will be absorbed by general-purpose agents, the development of model capabilities should be linked to economic benefits, and the true battlefield for agents lies in the real business scenarios of enterprises. To reach this level, deep synergy between industry and agents is required, and thus industrial agents came into being.
01.Enterprises Need Three Types of Agents Baidu AI Cloud Covers Them All
Shen Dou, Executive Vice President of Baidu Group and President of Baidu AI Cloud Group, presented a set of figures in his keynote speech: Tasks in enterprises are essentially divided into three categories, of which 20% are daily operation tasks, 30% are professional domain tasks, and 50% are core tasks that truly determine enterprise operations.
Corresponding to these three types of tasks, enterprises need three different agents, namely general-purpose agents, professional agents, and customized agents. General-purpose agents are responsible for basic work in the daily operations of enterprises, including organizing and aligning information, coordinating resources, etc. Baidu AI Cloud believes that the change brought by general-purpose agents is turning office capabilities into basic capabilities that can be called upon at any time. People can then shift more attention to professional tasks and core scenarios. DuMate is the product corresponding to this demand. The product has built a 3-layer Harness architecture and a 3-layer memory system, which can flexibly call various skills and resources, and can also remember users' usage preferences, becoming smoother the more it is used.
As of now, the monthly active users of DuMate have reached 6.743 million, a month-on-month increase of 1063.79%. In the past month, users grew by nearly 9 times, ranking first in growth rate among AI office agents. It has completed 150 iterations, and the enterprise version has released 15 suites and 96 skills. In the professional field, Baidu provides a professional agent matrix. FM Agent handles combinatorial optimization problems, Baidu Miaoda handles application development problems, Baidu Shengsuan handles data problems, Baidu Yijian handles visual problems, Baidu Hogee handles marketing problems, and Baidu Comate handles code problems.
As mentioned earlier, 50% of an enterprise's business consists of core tasks that reflect its unique processes, data, and know-how, which are the most important core businesses of the enterprise, and the agents used must be tailor-made. In response to such demands, Baidu AI Cloud has opened the same technology as Dazi to enterprises and launched the DuMate Open Platform. The platform provides enterprise-level agent development and operation services, with the same technological foundation, enterprise-level operation guarantee, and secure and controllable features, supporting flexible deployment such as public cloud, VPC, and privatization. At the same time, by building it into a fully open ecological platform, skills, suites, and applications can be more easily accessed and shared. Enterprises can use third-party capabilities through Dazi, and can also put their own capabilities on Dazi to serve more customers. Baidu AI Cloud will build a mutually beneficial open ecosystem around the implementation of industrial agents, and will not charge commissions from ecosystem partners.
Shen Dou emphasized in his speech that these three types of agents must not become new silos in the enterprise. What enterprises really need is to organize these agents into a system that can collaborate, close the loop, and be accountable for results around real tasks, forming an industrial agent operating system. This system takes on tasks upwards, determining which type of agent needs to participate; connects the enterprise's knowledge, data, permissions, tools, and processes in the middle; and schedules computing power, storage, and security resources on the cloud downwards. Simply put, it organizes agents, enterprise systems, and underlying cloud resources together, allowing agents to truly run within the business.
02.Used by 80% of Central SOEs and 100% of Mainstream Automakers Industrial Agents Begin to Bring Real Benefits
Can such a set of tools really bring tangible benefits to enterprises? Shen Dou revealed on-site that currently, 80% of central SOEs, 100% of mainstream automakers, and over 800 financial institutions and embodied intelligence enterprises nationwide are cooperating with Baidu AI Cloud to bring industrial agents into real production environments. For example, the insurance and financial services group Taikang has applied agents to core scenarios such as medical health, risk control and compliance, and sales operations. Take the "business compliance agent" as an example. In the past, business applications had to be reviewed item by item manually. The sources of rules were numerous, changes were fast, and the criteria were not easy to unify. Now, the agent first understands the key information in the application, matches the applicable rules, gives judgments and basis item by item, and then the reviewers review and confirm. In April this year, Taikang officially deployed Baidu AI Cloud's agent development platform. Up to now, more than 3,500 agents have been launched, with a daily average consumption of tokens reaching 20 billion.
China Southern Power Grid, through the data ontology capabilities of Baidu Shengsuan, organizes equipment, alarms, ledgers, maintenance records, and the business relationships between them. Combined with the AI workstation, it connects multiple business systems and has built more than a dozen agent applications for multiple core businesses.
Domestic leading whole vehicle R&D enterprise IAT worked with Baidu AI Cloud to integrate FM Agent into IAT's existing best CFD simulation system, allowing Famo to quickly generate and screen feasible aerodynamic drag design schemes, and then combine physical laws and high-precision simulation verification to continuously evolve better schemes. After 120 rounds of evolution, Famo found a better solution. Under the premise of ensuring the accuracy of aerodynamic drag prediction, a single aerodynamic drag evaluation has been shortened from several hours in the past to the second level, improving efficiency by hundreds of times.
These cases show that industrial agents have begun to bring real benefits, moving from marginal assistance to core business decision-making and production R&D.
03.Full-Stack Capabilities of Chips, Cloud, Models, and Applications Are Fully Supported Baidu AI Cloud Builds a Fortress from the Bottom Layer
For this operating system and agents to run, it cannot rely solely on the application layer; it must have underlying infrastructure and computing power support. The competition among agents is essentially AI cloud competition. This is not an exaggeration at all. According to predictions by the international data agency IDC, by 2030, the number of active agents globally will reach 2.216 billion, with a compound annual growth rate of 139%. Billions of agents will continuously call models, data, tools, and systems, bringing about over 100 gigawatt-level inference demands. This puts forward higher requirements for Agent Infrastructure (Agent Infra). Baidu AI Cloud's Agent Infra makes models completely resource-oriented: over 150 top models can be switched and called at any time. Over the past 6 months, through deep optimization at the inference engine level, the long-chain inference performance of its agents has improved by more than 3 times. At the same time, Baidu AI Cloud also provides Agent Cloud, which can achieve hot starts of tens of thousands of sandboxes in hundreds of milliseconds, unified management of agent identities and permissions, security guarantees covering the entire process from agent development to operation, full traceability of the agent execution chain, and real-time monitoring of core indicators for overall management and control.
At the bottom layer of Agent Infra, it relies on computing clusters with stronger performance and higher efficiency. To make industrial agents truly run, computing costs must continue to drop, and services must be stable enough. Kunlunxin has so far built fully domestic Kunlunxin clusters for multiple industries such as the Internet, energy, finance, manufacturing, automotive, operators, and education. Baidu Tianchi supernodes have completed thousand-cabinet-scale deployments in multiple intelligent computing centers across the country, successfully supporting the training of trillion-parameter large models, and serving industry customers in intelligent computing centers, finance, and the Internet. Today, Baidu Tianchi supernodes have achieved full-stack domestication. From core components to liquid cooling systems, from whole machine architecture to high-speed interconnection, a complete domestication path has been established.
In terms of gigawatt-level AI Data Center (AIDC), Baidu has turned power supply, cooling, IT, and monitoring into standardized prefabricated modules, completing integration testing in the factory and assembling them directly on-site, which can shorten the construction cycle by more than 30%. Through 800V DC power supply, phase change cooling, liquid cooling architecture, and source-grid-load-storage, its full-link power supply efficiency reaches over 98.5%, with an annual average PUE as low as 1.08, allowing every kilowatt-hour of electricity to release more computing power value.
This is Baidu AI Cloud's full-stack AI infrastructure: at the bottom layer are domestic computing power and cloud, turning electricity into usable, manageable, and sustainable tokens; moving up is Agent Infra, turning tokens into agents that can truly get things done; and further up, the three types of agents work collaboratively, which not only completes tasks one by one, but more importantly, leaves experience behind, turning it into capabilities that enterprises can reuse next time.
04.Focusing on Industrial Agents Baidu AI Cloud Has Started Its Flywheel
Why does Baidu AI Cloud focus on the track of industrial agents? Shen Dou's explanation is: Over the past three years, the biggest misjudgment in the industry was directly extrapolating the speed of economic growth from the growth rate of model capabilities. However, the reality is that model capabilities have been developing rapidly, but GDP has not been significantly boosted. The scissors gap in the middle is because model capabilities have not been cashed in on enterprise business scenarios. Lacking the combination of agents and industrial systems, they can only stay in marginal scenarios and fail to form an end-to-end closed loop. Therefore, Baidu AI Cloud believes that office is just one way for agents to enter enterprises. What truly requires professional delivery, customization, and is most valuable are the core scenarios of enterprise production and operation. What industrial agents bridge is precisely the gap between model capabilities and economic growth. According to Baidu's Q2 2026 financial report, Baidu's AI business revenue was CNY 12.5 billion, accounting for 50% of Baidu's general business revenue; AI cloud infrastructure revenue was CNY 7.3 billion, a year-on-year increase of 50%; GPU cloud revenue increased by 283% year-on-year.
Such impressive growth is the result of demand shifting from "training and inference computing power" to "Agent full-stack runtime". For Baidu AI Cloud, industrial agents are the second growth curve that is spinning up: The four layers of chips, cloud, models, and agents amplify each other, and the progress of each layer will bring greater returns to the other three layers.
05.Conclusion: Office Is Just the First Step Industrial Agents Are the Long-Term Main Battlefield
The AI industry is in a period of transformation. It took electricity thirty years to evolve from lighting to global infrastructure, and the Internet twenty years to evolve from portals to water, electricity, and coal. Today, AI is still in the first month after the invention of the electric light, and has not yet or has just begun to generate economic value. The true economic value of AI depends on whether it can enter the industrial system and become a part of the daily operation of enterprises. What industrial agents need to solve is precisely this "last mile": to make model capabilities no longer stay in Q&A and demonstrations, but to be embedded in the closed loop of production, R&D, and decision-making, bridging the gap between technology and economic growth.
Walking this path requires full-stack capabilities of chips, cloud, models, and applications, an open ecosystem for more partners to participate, and domestic computing power to take greater responsibility in terms of security compliance and controllable costs. Baidu AI Cloud's exploration is essentially building a set of infrastructure for industrial intelligence. As Shen Dou said, "Agents turn large models from a technology into a real economic force." And industrial agents are the most critical part in this process.