On 8 May 2026, the Cyberspace Administration of China (CAC), together with the National Development and Reform Commission (NDRC) and the Ministry of Industry and Information Technology (MIIT), published the “Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents” (智能体规范应用与创新发展实施意见). It is a policy document, not a law. It sets directions: the user’s rights over an agent’s decisions, control of behaviour, and risk-based classification with filing and recall of products in sensitive sectors. As of October 2026, China has no single law on AI.
The wording of the CAC document below is our working translation from Chinese.
In brief
- Definition: an agent is an intelligent system with capabilities of autonomous perception, memory, decision-making, interaction and execution.
- Three kinds of decisions (point 6): made only by the user, made on the user’s instruction, made by the agent on its own. The user has the right to know about the agent’s own decisions and the right to the final decision; the agent may not act beyond its authority.
- Control of behaviour (point 7): built-in rules and “behaviour guardrails”; for important scenarios, mechanisms that make the agent’s actions verifiable and traceable.
- Classification (point 11): in sensitive areas and key sectors, filing (备案), testing (检测) and recall of problematic products; in low-risk areas, self-assessment and self-regulation.
- Status: “implementation opinions” with no date of entry into force and no sanctions; carried out by the CAC, NDRC and MIIT.
Does China have a law on artificial intelligence?
As of October 2026, China has no single law on AI. Regulation is made up of separate acts:
| Act | What it regulates | In force from |
|---|---|---|
| Interim Measures for the Management of Generative AI Services | Public generative AI services | 15 August 2023 |
| Measures for Labelling AI-Generated Content | Labelling of synthetic text, images, audio and video | 1 September 2025 |
| Amendment to the Cybersecurity Law (adopted on 28 October 2025) | The first provision on AI in that law | 1 January 2026 |
| Interim Measures for the Management of Anthropomorphic AI Interaction Services (published on 10 April 2026) | Services that imitate human communication | 15 July 2026 |
| The Opinions on intelligent agents (CAC, NDRC, MIIT, 8 May 2026) | The application and development of AI agents | A policy document, with no date of entry into force |
The authorities announced in 2026 that a comprehensive AI law was being prepared; no official draft submitted to the National People’s Congress was found in public reports.
What do the Opinions say about AI agents?
The document has six sections and 38 points: principles, technological foundation, security, application by sector, ecosystem, and support. For the security of agents, the points of the section “Hold the bottom line of security” matter most:
- point 5: develop policy, legislation and ethical norms for agents; prevent abuse of data and of “humanization”, and protect minors and the elderly from addiction;
- point 6, “Define decision-making authority”: separate the decisions that only the user makes, the decisions that require the user’s instruction, and the agent’s own decisions, with the authority each of them needs; the user has the right to know about the agent’s own decisions and the right to the final decision; the agent’s operations must not go beyond the user’s instruction;
- point 7, “Strengthen behaviour control”: develop technologies of built-in rules and “behaviour guardrails”; study the use of blockchain and other technologies for mechanisms that make the behaviour of agents verifiable and traceable in important scenarios;
- point 8: the security of the agents themselves: management of permissions, control of behaviour, protection against data poisoning, leaks, substitution of algorithms and “getting out of control”;
- point 9: supply chain security: connection of models, API calls, extension tools;
- point 10: mechanisms for detecting risks, joint human and machine review, interception and blocking.
The section on the technological foundation (point 2) separately names tools for detecting anomalous behaviour and the ability to “detect, intervene, block and restore” when agents behave improperly; point 4 names a registration platform for agents with a digital identity and a declaration of capabilities.
How does the risk classification work?
Point 11 introduces differentiated management depending on the scenario and its possible impact:
- sensitive areas and key sectors: the internet authorities together with sector regulators define the admissible scenarios; filing, testing and recall of problematic products apply;
- low-risk areas (some scenarios of daily life, entertainment and everyday office work): self-assessment of compliance, disclosure, management through distribution platforms, and industry self-regulation.
Mandatory standards are proposed for medicine, transport, media and public safety (point 3).
What is not in the document?
- A date of entry into force and sanctions: these are “implementation opinions”, not a normative act.
- An explicit right to an instant forced stop of an agent: there is the ability to “detect, intervene, block, restore”.
- Extraterritorial effect: unlike the AI Act, the document does not apply directly to foreign providers.
- A formal scheme of several risk levels with mandatory filing at each: the division is into sensitive and low-risk areas.
The date of 15 July 2026, which is sometimes linked to this document, belongs to a different act: the measures on anthropomorphic AI interaction.
How does China’s approach differ from the EU’s and Russia’s?
| China (the 2026 Opinions) | EU (AI Act) | Russia (Law No. 243-FZ) | |
|---|---|---|---|
| Form | A policy document | A directly applicable regulation | A federal law |
| Agents named explicitly | Yes | No | No |
| Approach to risk | Sensitive sectors: filing and recall; the rest: self-assessment | Risk levels; high-risk systems under Annexes I and III | A risk-based approach as a principle; sector requirements from 1 March 2027 |
| Authority of the agent | Decisions only by the user, on the user’s instruction, and the agent’s own | Human oversight for high-risk systems | Not regulated |
More on the pages about the EU AI Act and about Russia’s Law No. 243-FZ.
For a company that builds agents, the main idea of the Chinese document matches the IS 3.0 architecture: every action of an agent must have an authority, going beyond it must be detected and stopped, and the behaviour must be verifiable and traceable.
The Russian version of this page also contains the institute’s May 2026 report on the implementation of China’s AI strategy; it is available in Russian only.
Frequently asked questions
Does China have a law on artificial intelligence?
There is no single AI law as of October 2026. Separate acts apply: the measures on generative AI (from 15 August 2023), on labelling AI content (from 1 September 2025), an amendment to the Cybersecurity Law with its first provision on AI (from 1 January 2026), the measures on anthropomorphic AI interaction (from 15 July 2026), and for agents, the CAC Opinions of 8 May 2026.
How does China regulate AI agents?
Through the “Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents” of 8 May 2026. The document requires separating the user’s own decisions, decisions made on the user’s instruction and the agent’s own decisions, forbids the agent to go beyond the user’s instruction, and introduces filing, testing and recall of problematic products in sensitive sectors.
Is the CAC document on agents binding?
It is a set of “implementation opinions”, a policy document with no date of entry into force and no sanctions. It sets directions for government bodies and industry; specific obligations will appear in standards and sector acts.
What are the “three kinds of decisions” of an agent?
Decisions that only the user makes; decisions that the agent makes on the user’s instruction; and the agent’s own decisions. For each kind, the limits and the authority must be defined; the user has the right to know about the agent’s own decisions, and the final decision stays with the user.
Sources
- CAC, “智能体规范应用与创新发展实施意见” (“Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents”), 8 May 2026 (in Chinese): https://www.cac.gov.cn/2026-05/08/c_1779979789523320.htm
- Caixin Global, “China amends Cybersecurity Law to add first-ever AI clause”, 29 October 2025: https://www.caixinglobal.com/2025-10-29/china-amends-cybersecurity-law-to-add-first-ever-ai-clause-102376988.html
Checked: 16 October 2026. The translation of the CAC document’s wording is the institute’s working translation.