A new set of policies has been released in Beijing to advance the development of intelligent agents, marking a significant push to lead in this emerging technology sector.
For the past six months, Beijing has focused on refining intelligent agents, progressing from foundational models to framework performance, inference efficiency, and industry applications. As a leader in China's AI development, Beijing has consistently taken the initiative. However, it also faces new challenges, including a shortage of high-performance computing power and the need for improved standards and service systems to support the large-scale deployment of intelligent agents.
To address these issues, the Beijing Municipal Development and Reform Commission, in collaboration with relevant city departments, released the "Several Measures of Beijing Municipality on Accelerating the Development of Intelligent Agents" on July 23. The document, consisting of ten articles, covers four key areas: technology, application, industry, and ecology. According to data from the China Business Industry Research Institute, the global AI agent market is estimated to be around $11.3 billion in 2025, a figure expected to reach $17.5 billion by 2026.
The Beijing Municipal Development and Reform Commission will dynamically reserve and implement a series of key projects through "open competition" and "horse racing" mechanisms. The selected projects will receive financial support of up to 100 million yuan. In the first half of 2026, the total financing scale in Beijing's artificial intelligence sector exceeded 95 billion yuan, accounting for over 30% of the national total, and the number of newly listed companies has already matched the total for the entire previous year.
Advancing Core Technology for Agent Control
Technology forms the foundation for the development of intelligent agents. The new measures focus on two key aspects. "On one hand, we are continuously pushing the boundaries of fundamental models," said Lin Jianhua, Deputy Secretary of the Party Leadership Group and Deputy Director of the Beijing Municipal Development and Reform Commission. He cited the "Touch the Sky" plan recently proposed by Tang Jie, founder of Zhipu AI, as an example. "This is also a consensus in Beijing. We hope that elites from all sides will share the ambition to continuously 'touch the sky,' supporting innovations in model architecture and training paradigms in areas such as online learning, continuous learning, and autonomous evolution, to constantly break through the limits of intelligence."
Lin Jianhua also emphasized the need to strengthen the development of common technologies for controlling intelligent agents. The measures specify support for implementing a control layer engineering project, focusing on areas like context engineering optimization, task persistence, multi-agent collaboration, and system scalability to solidify the common foundation for intelligent agents. This includes accelerating research on universal agent technology across different models and chips, improving the length of complex task chains and the stability of long-range task execution, and building a controllable, self-developed interconnection protocol, development framework, and basic tool chain. The goal is to develop standards and specifications, support the creation of "Ling Ji" and other agent infrastructure, enhance the deployment and cross-framework interoperability of intelligent agents, while ensuring security and controllability.
In 2025, the Municipal Bureau of Economics and Information Technology released the "Action Plan of Beijing Municipality on Supporting Information and Software Enterprises to Strengthen AI Application Service Capabilities (2025)" (the "AI Soft Eight Policy"). Now, the "Several Measures on Intelligent Agents" has been published. Chen Zhaohui, Deputy Director of the Beijing Municipal Bureau of Economics and Information Technology, provided a detailed explanation. "The AI Soft Eight Policy released in 2025 mainly addressed the issue of 'how enterprises can effectively use AI.' The current measures primarily solve the problem of how the industry can build a healthy ecosystem for intelligent agents."
"The policy upgrade is mainly reflected in three ways," Chen Zhaohui stated. "The policy positioning has shifted from AI empowering software to agents reshaping software. The support scope has expanded from software companies to the entire industry chain. And the policy system has evolved from point-based breakthroughs to a systematic reconstruction." Regarding the systematic reconstruction, the measures are comprehensive. For the first time, they introduce new concepts such as the intelligent agent infrastructure foundation, the Intelligent Agent Interconnection Protocol (AIP), the Token economy, and the OPC innovation and entrepreneurship model. This forms a policy system covering the entire lifecycle of an intelligent agent, from technological innovation and infrastructure construction to application scenarios, new business incubation, security governance, and factor guarantees.
Emphasis on Native Agent Applications
Regarding applications, the measures address two aspects: accelerating the deployment of native intelligent agent applications and promoting the integrated development of intelligent terminals and intelligent agents. For native agent applications, Lin Jianhua explained that the measures support the development of AI-native super software and artificial intelligence operating systems. This includes creating scientific AI assistants and platforms with capabilities for information correlation and knowledge system construction, building AI-native autonomous laboratories, and developing data-driven, autonomously deciding, and human-machine collaborative industry solutions. The measures also promote models like Frontier Deployment Engineers (FDE) to deepen industry application deployment, creating more benchmark demonstration scenarios in fields such as science, healthcare, and manufacturing.
As early as 2024, Beijing was the first in the country to release an "AI+" action plan. In the next step, relying on the deep evolution and widespread penetration of intelligent agents, the Beijing Municipal Development and Reform Commission will focus on breakthroughs in fields, regions, data loops, common platforms, and technologies. Lin Jianhua added, "We support the development of an AI OS with cross-device collaboration, cross-ecosystem development, and cross-scenario deployment capabilities, accelerating the application of intelligent agents in more terminal products. As early as last year, Beijing took the lead in launching the 'Doubao' phone. We will optimize inference engines to make intelligent agents more cost-effective and efficient. We will also build an agent development service platform to lower the barrier for developers."
Liu Weihua, a member of the Party Leadership Group and Deputy Director of the Beijing Municipal Science and Technology Commission and Zhongguancun Science Park Management Committee, provided a detailed account of the application of intelligent agents in the field of science and technology. "Scientific research agents can be understood as 'AI scientists.' Earlier, we continuously pushed for the upgrade and iteration of the Panshi science foundational model, just releasing version 2.0 last week. Using this as a base, we are building an integrated intelligent scientific research platform that allows researchers to quickly customize dedicated models and research agents on demand. Currently, multiple intelligent agents have been deployed on a large scale. For example, the Literature Compass agent can generate a professional-level review in three hours with 90% accuracy. It has generated nearly 40,000 reviews so far and provided over one million scientific question-answering inference services." Liu Weihua stated that in the next step, Beijing will accelerate core technology breakthroughs across the entire chain of scientific research agents. Using autonomous laboratories as core carriers, it will promote the efficient application of intelligent agents in scientific research, while coordinating the supply of computing power, data, and talent.
Supporting New Business Forms like OPC and the Token Economy
In the first half of 2026, new business models emerged in Beijing, including the minimalistic entrepreneurship model represented by OPC (AI One-Person Company) and the new business form represented by the Token economy. A number of industry companies achieved annualized revenue growth of several times or even more than ten times. The measures support entrepreneurial models like OPC as new forms of enterprise organization in the intelligent era. A hierarchical and categorized professional service system is to be built, upgrading public services around flexible computing power supply and professional incubation. To meet the needs of OPCs regarding market access, pilot explorations will be actively carried out to make OPCs easier to establish and grow.
The measures also clarify the need to closely follow the inherent needs of the Token economy. This includes supporting the development of dedicated inference chips, conducting key technology research in inference architecture, and optimizing application-layer efficiency through inference cache reuse and intelligent task routing to improve Token efficiency across the entire chain. New business models such as Token-as-a-Service (TaaS), Agent-as-a-Service (AaaS), and Result-as-a-Service (RaaS) will be fostered. More Token products and services will be included in the scope of service vouchers for small and medium-sized enterprises. At the same time, the transition of Token billing from a focus on quantity to a focus on quality will be encouraged to improve economic efficiency and quality, and the monitoring and evaluation methodology for the Token economy will be perfected.
Taking the AI Origin Community as an example, Tang Chao, Deputy Director of the Zhongguancun Science City Management Committee and Vice Mayor of Haidian District, said, "The AI Origin Community has gathered over 1,000 AI scientists, more than 13,000 AI developers, and over 100,000 young AI students. It serves more than 30 surrounding universities and research institutions and has attracted over 400 AI companies, forming a '10-minute innovation circle.'" Tang Chao highlighted achievements in concentrating source innovation resources, connecting the entire OPC training chain, and providing diverse supporting services. "Haidian District has set up exclusive open workspaces for OPCs, launched the 'OPC Order Taking' platform to supply application scenarios, and has gathered nearly 180 OPC companies, primarily young people aged 25-30. Haidian has introduced a combined policy of computing power subsidies and Token vouchers, reducing the cost of computing power calls. Simultaneously, it has established a flexible computing power lab, collaborating with cloud service companies to build an inclusive computing power supply system."
The Economic and Technological Development Zone also has its own plans and achievements for OPC growth and collaboration. According to Li Yantao, Director of the Beijing Economic and Technological Development Zone Management Committee and Director of the Beijing Municipal Integrated Circuit Major Project Office, the zone has built 8 OPC communities, including the city's first. It released the first domestic open-source common infrastructure for intelligent agents, providing a secure and controllable operating base for OPCs. In cooperation with leading domestic AI open-source communities, the zone has established a full-cycle growth service system for OPCs, offering one-stop services from technical exchanges and resource matching to business implementation. An innovative OPC business collaboration mechanism has been built, and a super-individual intelligent task distribution platform has been set up, facilitating the landing of over a hundred high-quality orders.
Preventing Malicious Abuse of Technology
Security governance is a guarantee for the development of intelligent agents. The measures explicitly call for balancing security governance with open-source development. This involves improving the governance capabilities of intelligent agents, further regulating the industry to prevent malicious technology abuse, developing security models to enhance the ability to "govern models with models," and accelerating the improvement of common service capabilities such as agent vulnerability scanning, security detection, attack and defense drills, and security certification.
Li Degang, Deputy Director and Spokesperson of the Beijing Municipal Internet Information Office, detailed the construction of an innovative, open, inclusive, and prudent governance model. "We must leverage Beijing's advantages in technology, talent, and 'Two Zones' policy. Based on the already-approved smart assistants from Xiaomi, Samsung, and the Doubao phone by the Central Cyberspace Affairs Commission, we will focus on business explorations such as the integration of software and hardware for intelligent agents and their value assessment. We will fully implement the management policy for marking AI-generated synthetic content, implementing both 'explicit' and 'implicit' labeling requirements to standardize the healthy interaction of AI-generated content. We will also effectively implement the 'Interim Measures for the Management of AI Anthropomorphic Interactive Services.'"
For key elements like talent, computing power, and capital, the measures also propose specific support plans. According to the latest information shared by Li Yantao, the Beijing Economic and Technological Development Zone has taken the lead in releasing an OPC talent recognition standard. Qualified OPCs will be included in the Yicheng talent system, with supporting social security continuation and insurance safety nets during the entrepreneurial transition period, providing worry-free support for OPC entrepreneurs. The zone also issued the nation's first virtual idol identity certification. It has innovated a sandbox approval mechanism, allowing pilot enterprises to conduct digital human commercial performances within the sandbox, thereby accelerating the commercialization of new business forms.
Comments