CAICT and CUPL Release Agent Governance Research Report (2026)
CAICT and China University of Political Science and Law released the Agent Governance Research Report (2026). The report sets a three-layer governance frame covering foundation, operation, and ecosystem layers.
In 2026, agent systems moved further from conversational response toward task execution through perception, planning, and tool use. The China Academy of Information and Communications Technology (CAICT) Policy and Economic Research Institute and the Artificial Intelligence Law Institute of China University of Political Science and Law (CUPL) released the Agent Governance Research Report (2026).
The report reviews 2026 technical progress and national governance practice. It describes new conditions in an agent era, maps AI governance questions raised by autonomy, interaction, and learning, and proposes a three-layer governance frame covering foundation, operation, and ecosystem layers. The authors state the aim of aligning compliant use with continued innovation.
Core findings
First, agents are entering a new development stage and are a contested segment of the AI industry. Benchmark scores have risen, and agents are shifting from passive dialogue boxes toward systems that act, while long-horizon complex tasks remain weak. Market competition is forming around traffic entry points, user lock-in, and commercial closed loops, with on-device agents as a contested product class. Agents are also changing interaction toward more proactive modes, networks toward more collaborative forms, and economic activity toward more automated operation.
Second, the risk map shows stepwise transmission and new governance questions. Risks associated with large models migrate and grow: hallucination can move from incorrect text to incorrect action, and privacy exposure can increase. Autonomy adds unauthorized operation and tool misuse. Interaction and collaboration add cascade failure and gaps in responsibility tracing. Effects extend to labor markets, monopoly and unfair competition, and financial stability. The report also flags erosion of human autonomy and loss of control at the frontier as outer bounds.
Third, major economies are taking different paths. The United States has a frontier-model review frame, with state law and enforcement filling responsibility and transparency rules. The European Union uses fundamental-rights protection and the EU Artificial Intelligence Act risk classification across the agent value chain. Singapore has published what the report calls the first agentic AI governance framework, as part of a trusted-AI hub policy. China, responding to AI Plus application demand, has issued agent-specific rules for a safety baseline, is using standards to shape agent interconnection, and is exploring governance that stacks new agent rules on existing large-model rules.
Fourth, the report proposes governance in three layers. The foundation layer covers data and identity: identity management as a trust starting point, least privilege and dynamic authorization as action bounds, and isolation and encryption of long-term memory for data management. The operation layer covers execution and dynamic control: review and assessment, transparency disclosures, dynamic monitoring, and intervention. The ecosystem layer covers value coordination: standardized communication protocols and interface security for trusted interconnection, shared capability modules and toolchains, and tracing plus multi-agent responsibility assignment.
Fifth, future work is grouped in four tracks: a dual emphasis on safety and development with a controllable baseline; institutional rules on property, liability, and authority boundaries; technical work on identity and permission management, safety testing and evaluation, and risk monitoring and intervention; and ecosystem work on open interconnection, fair competition, international coordination, and multi-party governance.
Report contents
Chapter one covers new conditions: the shift from environment sensing to autonomous action, faster market-structure formation, and paradigm change in the intelligence transition.
Chapter two covers new AI governance questions: the agent risk map and the logic of agent governance.
Chapter three covers national trends: US frontier-capability diffusion controls with model review and standardization; EU extension of existing rules to new scenarios; Singapore multi-dimensional governance for a trusted AI hub; and China stacked large-model and agent rules under AI Plus demand.
Chapter four details the three-layer frame. Chapter five outlines the four future tracks listed above.
Authors, copyright, and access
Named CAICT authors include Li Qiangzhi, deputy director of the Policy and Economic Research Institute and professor-level senior engineer; Cheng Ying, chief engineer of the institute’s supervision research department; Liu Zhipeng, senior engineer; and Zhang Meng, engineer. Additional CAICT contributors listed are Gao Yongchuang, He Jia, and Han Chenyang.
Cheng Ying is listed as the writing-team contact at chengying@caict.ac.cn.
The report is copyrighted by CAICT and China University of Political Science and Law. The publishers require that reuse of text or views credit both institutions.
Download: https://www.caict.ac.cn/kxyj/qwfb/ztbg/202609/P020260916574260931247.pdf