On 8 May 2026 the Cyberspace Administration of China, the National Development and Reform Commission and the Ministry of Industry and Information Technology jointly issued the Opinion on the Standardised Application and Innovative Development of AI Agents. The document implements the State Council’s Opinion on Deepening the “AI Plus” Initiative and upholds the principles of controllable security, orderly standardisation, innovation-driven development and application-led progress.
An “AI agent” is an AI system able to perceive, plan and execute tasks autonomously. Unlike a model that only answers questions, an agent can call tools, access systems and chain multiple steps together. This shifts the focus of governance from “what the model outputs” to “what the agent does”.
From model governance to behaviour governance
Previously the governance of generative AI centred mainly on the model and content layers: filing, labelling and content security. AI agents change that. When a system can execute operations autonomously, call external interfaces and complete tasks on a user’s behalf, the risk is no longer only “what it says” but “what it does”.
This Opinion places data security, personal-information protection, permission management and behaviour control side by side, matching that shift: permission management addresses “what it can move”, behaviour control addresses “how it does it”, and data and personal-information protection address “what data it can touch”. Together they form a governance framework for agent behaviour; see our AI agent interconnection national standards analysis.
Nineteen typical application scenarios
The document sets out nineteen typical application scenarios, covering fields such as education and healthcare. The list grounds the abstract “agent” in specific industries: education involves protection of minors and content suitability; healthcare involves the handling of personal information, including sensitive personal information, and the boundaries of responsibility. When choosing an entry point, companies must assess both technical fit and the scenario’s compliance requirements. Directions included among the typical scenarios also tend to attract pilots, standards and supporting measures.
Tiered governance and compliance self-assessment
The document proposes promoting the standardised application of low-risk scenarios through compliance self-assessment, information reporting, distribution-platform management and industry self-discipline, pointing to tiered governance: low-risk scenarios emphasise standardised application, where companies can make a preliminary judgement through self-assessment; high-risk scenarios require more rigorous assessment and reporting. Self-assessment means companies must build an operable internal review capability rather than wait for external inspection, while distribution-platform management extends responsibility to the channel layer, where platforms must carry corresponding duties for the capabilities they list; see our Algorithm Filing service.
Supply-chain security and permission management
The document proposes exploring a supply-chain security information-sharing and early-warning mechanism for AI agents. This targets the componentised nature of agents — an agent is often made up of models, tools, plugins and external interfaces, and a problem in any link can transmit along the chain. Companies therefore need finer supplier and component management: the security and compliance status of the models, tools and third-party services used must enter the assessment, and liability boundaries must be clear in contracts.
Permission management is another specific requirement. An agent gaining the ability to act presupposes a clear scope of authorisation, auditable operation records and an intervene-able control mechanism; without these, over-reach or mis-operation is hard to attribute. Overseas regulators are watching agent risk too; see our ICO agentic AI compliance guide and AI safety governance framework analysis.
What it means for companies
The document builds out the governance framework for agents: dividing risk by scenario, configuring requirements by tier, implementing responsibility through self-assessment and reporting, and covering the chain through supply-chain security. The pragmatic pointers are to define permission boundaries and data scope at the product-definition stage, assess compliance requirements alongside scenario selection, and bring security and compliance status into supplier management when procuring components.
For support on data compliance, permission design, algorithm filing and supply-chain security arrangements for AI agents, see our Data Compliance service.
This article is based on public information on the Opinion on the Standardised Application and Innovative Development of AI Agents jointly issued by the Cyberspace Administration of China, the National Development and Reform Commission and the Ministry of Industry and Information Technology on 8 May 2026. It is for reference only and does not constitute legal advice. Specific implementation requirements should be based on the official document.