This page summarizes the English translation of the Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents, issued jointly on 8 May 2026 by the Cyberspace Administration of China (CAC), the National Development and Reform Commission (NDRC) and the Ministry of Industry and Information Technology (MIIT). The document is a non-binding policy instrument directed specifically at agentic systems. The translation cited here was published by Geopolitechs on the day of issuance (Source: geopolitechs.org). See the legislation page for structured provisions and comparisons.
Provenance
- Issuers: Cyberspace Administration of China, National Development and Reform Commission, Ministry of Industry and Information Technology
- Date of issuance: 2026-05-08
- Translation: Geopolitechs, geopolitechs.org
- Chinese original: mp.weixin.qq.com (not retrieved)
- Accompanying material: an official five-question Q&A with the press, published via Xinhua and reproduced in the same translation
- Stated legal basis: the Opinions state that they implement the State Council's Opinions on Deepening the "AI+" Initiative
The English rendering is a secondary translation of a primary instrument. Provisions turning on exact wording should be checked against the Chinese original.
Structure
The document opens with a definition — agents are "intelligent systems capable of autonomous perception, memory, decision-making, interaction, and execution" — and is organized into six parts: fundamental principles; strengthening the foundations for development; safeguarding security baselines; strengthening application-driven development; building an innovation ecosystem; and safeguard measures. Thirty-eight numbered tasks run continuously across the parts.
Fundamental principles
Part I states four principles. Safety and controllability treats the safety, reliability and trustworthiness of agents as baseline requirements "throughout the entire lifecycle of agent technology R&D, deployment, and commercialization," with the aim of preventing systemic risks. Standardized and orderly development calls for a governance framework aligned with existing law that encourages industry self-governance and "clearly defines bottom lines and red lines." Innovation-driven development couples theoretical, technological and engineering innovation and calls for collaborative mechanisms among government, industry, academia, research institutions and users. Application-driven development directs effort at practical needs in scientific research, industrial development, consumption promotion, public welfare and social governance, promoting "gradual, step-by-step implementation."
Technological and standards foundations
Part II addresses the development base. On technology, it calls for continued improvement of general-purpose foundation models alongside domain-specific models "adaptable to different scenarios and devices," greater supply of high-quality training and operating datasets, and research on task understanding, task planning, tool usage, long-term memory, interoperability and multi-agent collaboration. It also calls for agent toolchains covering development, testing, deployment and operations, and for security and governance tools such as adversarial-sample detection and abnormal-behavior detection supporting discovery, intervention, blocking and recovery from non-compliant agent behavior.
On standards, the Opinions call for an overall standards framework covering key technologies, major products, data exchange, application scenarios, quality evaluation, security assurance and trustworthy certification, and for foundational standards governing interfaces between agents and software tools, application services and hardware peripherals. They name agent interoperability protocols (AIP) for promotion as national and industry standards, and support mandatory standards in healthcare, transportation, media and public security. Participation in international standards development is encouraged.
The Intelligent Internet
Task 4 sets out an "intelligent internet" programme with no direct counterpart in United States or European Union instruments. It calls for study of an intelligent internet system architecture and exploration of agent registration platforms providing digital identity management, discovery and search, and capability declarations for agents, supporting inquiry and management of developers, deployment methods, interface protocols and compliance certifications. Supporting research areas named are agent identity systems, trusted interconnection, compliant payments, security protection and conflict resolution. The Opinions direct that the advantages of IPv6 be used to improve end-to-end agent communication, and call for monitoring and evaluation indicators for the intelligent internet.
Security and governance provisions
Part III is framed as "people-centered and AI-for-good," featuring "multi-stakeholder governance and prudent security management."
Decision-making authority (task 6). The Opinions direct that the boundaries and required permissions be clearly defined for three categories: decisions reserved exclusively for users, decisions requiring user authorization, and autonomous agent decisions. Users are to retain "the right to know and the ultimate decision-making authority over autonomous agent actions," and agents are not to exceed the scope of user authorization.
Ethical norms and anthropomorphism (task 5). Agent behavior is to comply with law and "mainstream societal values." The text directs that agents be prevented from "exploiting data advantages or anthropomorphic technologies to spread harmful values or engage in algorithmic exploitation," and specifically names risks of "addiction and emotional dependence among minors and the elderly."
Behavioral controls (task 7). Embedded rules and behavioral guardrails are to keep agents operating lawfully in public, private and specialized environments. The Opinions call for exploring blockchain and related technologies to establish verifiable and traceable mechanisms for agent behavior in key application scenarios.
Security risks (tasks 8–10). Named intrinsic-security research areas are data security, personal information protection, cryptographic safeguards, attack detection, permission management and behavioral controls, guarding against data poisoning, privacy leakage, algorithm tampering, system vulnerabilities and loss of operational control. Supply-chain provisions cover model access, API calls and extension-tool use, with information-sharing and early-warning mechanisms and timely risk alerts. Application-derived risk provisions call for human-machine collaborative review, interception and blocking, and for preventing agents being used for automated attacks, privacy violations, false information generation and dissemination, and online fraud.
Tiered and classified governance (task 11). Governance is keyed to application scenario and potential impact. For sensitive sectors and key industries, cyberspace authorities together with relevant regulators are to determine permissible application scenarios and apply filing requirements, testing and product recalls. For lower-risk fields such as "daily entertainment and office work," governance runs through compliance self-assessment, information reporting, platform governance and industry self-regulation.
Compliance services and self-regulation (tasks 12–14). The Opinions call for third-party testing and evaluation of agent functionality, performance, quality and compliance, mutual recognition of certification results, and published reports on agent technological and application maturity. Industry associations and major enterprises are encouraged to formulate self-regulatory rules covering functionality compliance, algorithm governance, intellectual property protection and fair competition. Task 14 directs industry organizations to establish voluntary credit evaluation mechanisms for agent-ecosystem market participants, with "lawful disciplinary actions" against technology abuse, induced consumption, false advertising and concealment of defect information; developers, development platforms, distribution platforms and service providers are all named as participants.
Application scenarios
Part IV sets out typical application scenarios across five domains, which the official Q&A counts as nineteen. Scientific research covers research exploration and R&D assistance, including software engineering agents and integration with CAD and CAE tools. Industrial development covers intelligent manufacturing, energy and resources, transportation, agricultural production and financial services. Consumption covers terminal applications, culture and tourism, and commercial services, the last including embodied agents for guidance, cleaning, warehousing and distribution and for "low-cost domestic services, elderly care, childcare, and disability assistance." Public welfare covers education, healthcare, human resources and information services. Social governance covers government services, judicial services, public safety, urban governance, and tendering and bidding.
Two scenario provisions bear on governance debates elsewhere in the document. Task 28 (information services) encourages content platforms to develop agents for "intelligent moderation, public opinion guidance, emotional counseling, and real-time translation." Task 30 (judicial services) contemplates full-process judicial assistance agents for organizing case materials, entering case information, reviewing evidence and generating legal documents.
Innovation ecosystem and implementation
Part V calls for open-source cultivation, including compatibility between agents and open-source chips, operating systems and foundation models; industrial collaboration platforms such as agent ecosystem alliances and technology validation laboratories; agent software stores and supply-demand information platforms; opening of key application scenarios through pilot programmes; and cultivation of a global ecosystem using the World Artificial Intelligence Conference and World Internet Conference, with support for overseas compliance and localization by Chinese firms.
Part VI assigns coordination to the CAC, NDRC and MIIT together with relevant authorities, and calls for a comprehensive evaluation indicator system for agent development with monitoring, evaluation, rolling implementation and dynamic adjustment mechanisms.
Official Q&A
The accompanying press Q&A states that agent products "represented by mobile assistants, terminal-based intelligent managers, and cloud-based agents have emerged rapidly and are being applied at scale," while "the high autonomy and high permission levels of agents have also created security risks such as privacy leakage, unauthorized operations, and loss of behavioral control."
It situates the Opinions as implementing the State Council's Opinions on Deepening the Implementation of the "AI+" Initiative, issued in August 2025, which set a phased target that by 2027 AI would achieve broad and deep integration with key sectors and that "the adoption rate of next-generation intelligent terminals, agents, and other applications would exceed 70%." The Q&A cites Xi Jinping's remarks at the 20th group study session of the Political Bureau of the CPC Central Committee on accelerating relevant laws, regulations, policy systems, application norms and ethical guidelines, and building systems for technological monitoring, risk warning and emergency response.
The Q&A describes the drafting emphasis as fourfold: ideological guidance, coordinating development and security, application-driven development, and safeguarding the security baseline. It summarizes the regulatory requirements as clarifying product principles, preventing security risks, improving the governance system and strengthening industry self-regulation, with the stated aim that agents be both "allowed to develop dynamically" and "well regulated."
Relationships
- instance-of: Agentic AI — the class of systems the Opinions govern.
- related: China — Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents — the structured legislation page for this instrument.
- related: Cyberspace Administration of China (CAC) — lead co-issuer.
- related: China — Interim Measures for the Administration of AI Anthropomorphic Interaction Services — the binding instrument addressing anthropomorphic interaction and emotional dependence.
- related: Human Oversight — the three-tier decision-authority framework.
- related: China — Interim Measures for the Management of Generative AI Services, China — Provisions on the Administration of Deep Synthesis Internet Information Services, China — Internet Information Service Algorithmic Recommendation Management Provisions — the surrounding regulatory stack.