The exhibitors demonstrated the multi-dimensional holographic AI sensory behavior research system at the 5th China International Consumer Products Expo. Photo by Xinhua News Agency reporter Guo Cheng

  Since the beginning of this year, from the central government to the local government, a series of deployments have been made to cultivate embodied intelligence and develop intelligent robots. With the rapid development of technology, robots are no longer far away when entering your and my life. Starting today, our newspaper will launch a series of reports on "Robots are by your side" to focus on the development and changes of the robot industry.

  With the continuous breakthrough of large models, the popularity of AI agents continues to increase. Since the beginning of this year, from start-ups to technology giants, they have accelerated their layout, and intelligent application scenarios have continued to expand. Beijing, Shanghai and other places have successively issued policies to inject new impetus into the development of intelligent bodies. What is an AI agent and what can it do? Why are AI companies deploying intelligent bodies? What issues should be paid attention to when developing an agent?

  Multi-subject layout intelligent body

  It is generally believed that an AI agent is an advanced artificial intelligence system that can automatically perceive, think and act in a specific environment. It can understand, learn and reason to perform complex tasks and make decisions. At present, intelligent bodies have been applied in content creation assistants, knowledge question and answer assistants, intelligent assistants, AI search and other scenarios.

  "In a broad sense, AI agents are an intelligent application that can think deeply, plan independently, make decisions and execute deeply, but in the process of implementation, each manufacturer and product will undergo different changes and combinations according to user population and usage scenarios." said a relevant person in charge of Quark, an AI application under Alibaba.

  Since the beginning of this year, various companies have successively launched intelligent products that gather their respective advantages and expertise. The domestic big model team Monica released the general AI intelligent product Manus and announced that it had officially reached a strategic cooperation with the Alibaba Tongyi Qianwen team. Zhipu AI releases AI agent product AutoGLM to contemplate, pushing AI agents into a new stage of "thinking while doing". Quark creates a super intelligent body in the form of "AI super box" and has advantages in search, browser, scanning, and question shooting. Lenovo Group has successively launched "urban super intelligent bodies" in cities such as Wuyishan and Yichang, providing customized "artificial intelligence +" engines for thousands of industries and users. ByteDance Agent's product "button space" has recently started internal testing.

  What can current agents do? Li Wei, deputy director of the Institute of Cloud Computing and Big Data of the China Academy of Information and Communications Technology, said that it seems that in terms of enterprise-level applications, it is possible to use agents to promote further practices of enterprise cost reduction and efficiency improvement and data-driven decision-making. For example, in the logistics industry, intelligent bodies combine warehousing robots to realize automatic sorting and path planning, reduce sorting error rates, and improve warehouse throughput and flow efficiency; in the field of human resource management, intelligent interviewers deployed in the cloud can elastically expand and process massive video interview data, and use cloud GPU-accelerated multimodal analysis to evaluate candidate matching, which not only increases the efficiency of HR initial screening by 80%, but also continuously optimizes the evaluation model through the cloud continuous learning mechanism to avoid human bias.

  In terms of consumer-level applications, Li Wei believes that the agent focuses on providing scenario-based services and personalized experiences to individuals. For example, the intelligent marketing engine of the fast-moving consumer goods industry runs in the cloud, and dynamically generates recommendation strategies and publicity contents for thousands of people through real-time analysis of user behavior data flow; the cloud-edge collaborative intelligent body in the smart home scene uploads user habitual data to the cloud to train personalized models, and then controls home environment equipment in real time through edge computing nodes to automatically optimize the home environment in temperature, lighting, ventilation and other aspects.

  In short, the development of an agent will reshape the relationship between people, AI tools, and tasks. Wu Jia, vice president of Alibaba Group, said recently that many people say that in the AI ​​era, we should re-made all products. In fact, the real meaning is that in order to make AI better use of tools, we should re-made these tools, "AI uses tools, and people use AI."

  Evolve to commercialization

  Driven by the two major factors of technological iteration and market demand upgrading, startups and Internet manufacturers have successively deployed intelligent products and intelligent development platforms. The two-way resonance between underlying technological breakthroughs and industrial value release is catalyzing the acceleration of the evolution of agents from technological concepts to commercialization. According to CIDI Consultant's estimate, the global AI agent market size will continue to grow at an average annual compound growth rate of more than 40% in the next five years.

  Policy winds accelerate the industrial process. On April 8, the Beijing Municipal Bureau of Economics and Information Technology issued the "Beijing Action Plan for Supporting Information Software Enterprises to Strengthen the Service Capacity of Artificial Intelligence Applications (2025)", proposing to support the development of general agents and support innovative entities to develop cross-domain, multi-tasking, self-planned general agents. For general agents that have obtained batch numbers for generative artificial intelligence products and services and are listed on various application stores for the first time, priority will be given to the coordinating computing power guarantee, and the cost of calling computing power and model in the operation services will be provided with up to 30 million yuan.

  On April 21, the Shanghai Municipal Economic and Information Technology Commission issued the "Notice on the Revealing of the New Generation of General Artificial Intelligence Innovation Tasks in 2025", which mentioned that it would explore many agents with organic synergies and heterogeneity in complex open environments, and achieve sustainable emergence of group intelligence, and address key technologies such as multi-agent system and optimization decision-making technology, unmanned cluster system technology, group intelligence software technology and group intelligence federated learning technology.

  In terms of personal applications, Quark uses its own technical capabilities to help achieve users' needs in a wide range of scenarios. "Quak uses a minimalist 'AI super box' to be a personal all-round assistant." Wu Jia said that in the future, humans no longer need to directly use search and other tools, but instead hand over the complete task instructions directly to AI, which will think and execute and complete the final task delivery. Smartphones also understand humans better with the support of their intelligent bodies. It is reported that Honor YOYO agents can currently complete 600 requirements and intention understanding, 950 personal habit memory, and 270 complex task planning. Order coffee, query or cancel automatic renewal, and generate ID photos can all be achieved in one sentence.

  In terms of empowering enterprises, Volcano Engine has provided a variety of AI agent solutions to customers in multiple industries through the HiAgent platform. "In the financial industry, the intelligent body makes relevant attempts and explorations in scenarios such as customer service, wealth management, and industry development assistants; in the medical industry, the intelligent body makes relevant attempts and explorations in scenarios such as intelligent guidance, in-hospital services, and patient health management; in the education industry, the intelligent body makes relevant attempts and explorations in scenarios such as campus Pepsi, teacher courseware assistants, and scientific research assistants; in the manufacturing industry, the intelligent body makes relevant attempts and explorations in scenarios such as equipment maintenance assistants, R&D assistants, and embodied intelligence." A relevant person in charge of Volcano Engine introduced.

  In the field of smart cities, Lenovo Group recently launched the "urban super intelligent body" in cities such as Wuyishan and Yichang, adopting the "1×N intelligent body solution", that is, a super intelligent body and intelligent body in multiple fields work together to deliver AI capabilities to all walks of life and citizens and tourists in the city, realizing the comprehensive intelligence from government affairs to people's livelihood and industries, forming large scenarios with mixed small scenarios, and promoting the intelligent development of the city. "The agent not only retains the powerful data processing capabilities of the big model through the 'playing strengths and making up for the weaknesses' method, but also introduces self-boundary judgment, active perception, complex task decomposition and memory mechanisms, making artificial intelligence closer to the practical application scenarios of human intelligence." said Dai Wei, senior vice president of Lenovo Group and general manager of China Solution Service Business Group.

  Han Zizhe, an analyst at the Big Data and Artificial Intelligence Industry Research Center of CIDI Consulting, said that the independent decision-making ability and tool calling ability of the agent make AI no longer an isolated technical module, but a "productivity unit" that can be deeply embedded in the enterprise operation system, which is expected to solve the problems of fragmentation of traditional AI application scenarios and low input-output ratios, which will promote the implementation of AI technology on the industry side.

  Focus on various challenges and risks

  As the industry surges toward agents, technical challenges and multiple risks follow.

  The first thing to be wary of is the "pseudo-agent". Li Wei said that with the rise in popularity of the concept of agents, a group of "pseudo-instruments" under the banner of agents have emerged in the market. Many companies have packaged traditional technologies and existing products into agents, and promoted them through marketing strategies to mislead users.

  Secondly, we need to pay attention to technical bottlenecks. Han Zizhe said that we should pay attention to the technical bottlenecks in the cognitive reliability of the agent, including machine illusion, decision-making black boxes and other issues, which is the technical basis for ensuring the agent's trustworthy decisions. Li Wei said that AI agents rely highly on large language models, but their accuracy of generated content and contextual understanding capabilities still cannot meet the application needs of production-level scenarios. The existence of machine hallucinations may lead to decision-making errors in high-risk areas such as finance and medical care.

  Again, the cost issue also needs attention. Token consumption increases significantly when using agents to perform large and complex tasks. For example, Manus controls the inference cost to about one-tenth of the inference cost of DeepReaserch, an agent owned by OpenAI, and each task still needs to consume 500,000 to 2 million tokens. "With the increase in user volume and user usage, the exponential growth of computing power resources is an inevitable cost investment for intelligent bodies to implement. The computing power demand of super-large-scale GPU clusters far exceeds the scheduling capabilities of traditional data centers. It is necessary to rely on new computing power interconnection scheduling architectures such as AI cloud and computing power Internet to achieve efficient and low-cost computing power calls." Li Wei said.

  Finally, we should pay attention to the issue of technical ecology and collaborative standardization. The issue of technical ecology and collaboration standardization is the key to determining the scale implementation and cross-scenario collaboration capabilities of AI agents. However, the current protocol standard "match between multiple strong" has not yet formed a unified standard, and the coordination efficiency has not been effectively improved. Han Zizhe believes that industry users need to pay attention to the construction of internal high-quality data sets, accumulate data based on business scenarios, drive the optimization of the agent with high-quality data, and then feed back the business value to improve, thus forming a benign closed loop. Li Wei suggested that the industry urgently needs to build unified standards similar to Internet protocol clusters, optimize multi-agent architecture to release synergistic potential, and jointly promote AI agents to move from "single-point breakthrough" to "ecological prosperity."

  From technological verification to a critical period of large-scale implementation, the development of AI agents is highly consistent with my country's economic and social digital transformation needs. Looking to the future, we need to coordinate technological innovation and the construction of the governance system. While consolidating the foundation of independent and controllable technology, we will promote the coordinated research of industry, education, research and application, explore a sustainable model of deep integration of intelligent technology and the real economy, and inject lasting momentum into the leap in productivity in the digital age. (China Economic Net reporter Li Fang)

[Editor in charge: Ran Xiaoning]