Implementation Opinion on the Standardised Application and Innovative Development of AI Agents
Status and character
Section titled “Status and character”Published on 8 May 2026, the Implementation Opinion on the Standardised Application and Innovative Development of AI Agents is a policy and implementation document, not a statute or departmental rule. It operationalises the State Council’s “AI+” programme for systems capable of perception, memory, decision-making, interaction and execution.
Main governance directions
Section titled “Main governance directions”- build standards for agent technologies, products, data exchange, evaluation, safety and certification;
- study agent registration, digital identity, capability declarations and trusted interoperability;
- distinguish decisions reserved to users, decisions requiring user authorisation and decisions agents may make autonomously;
- preserve user knowledge and final decision rights and prevent actions beyond granted authority;
- develop behaviour controls, traceability, security testing and supply-chain risk sharing;
- address addiction, emotional dependency, privacy, data poisoning, privilege abuse and loss-of-control risks.
Governance architecture
Section titled “Governance architecture”The Opinion separates agent governance into four operational layers:
- Permission and behaviour controls. Decisions are divided among user-only decisions, decisions requiring user authorisation and decisions that an agent may take autonomously. Users retain a right to know and a final-decision right; execution may not exceed the granted authority.
- Security by lifecycle and supply chain. The document calls for controls covering development, deployment, operation and maintenance, including model access, API calls and extension tools. It specifically identifies poisoning, privacy leakage, algorithm tampering, vulnerabilities and loss of control.
- Classification by context and impact. Sensitive sectors and key industries may use filing, testing and product-recall measures. Lower-risk entertainment and office uses may instead rely on self-assessment, information reporting, app-distribution controls and industry self-regulation.
- Assurance infrastructure. Third-party testing, certification, monitoring, credit evaluation and interoperable assessment results are encouraged, but the Opinion itself creates no certification licence.
Application and implementation layer
Section titled “Application and implementation layer”The document identifies 19 application directions spanning scientific research, manufacturing, energy, transport, agriculture, finance, consumer devices, tourism, commerce, education, healthcare, employment, information services and public administration. Of particular governance significance are proposed agents for administrative approvals, judicial assistance, public safety, urban management and public procurement. These examples are policy priorities, not independent authorisations to automate legally reserved decisions.
The Opinion therefore sits between the State Council’s broad “AI+” strategy and later sector-specific rules, standards and pilots. A concrete local implementation example appeared in Shanghai in June 2026, where a trial measure for a government-procurement supervisory agent preserved human verification and decision responsibility.
The document is important because it moves Chinese AI governance beyond content-generation services toward tool-using and action-taking systems, while leaving binding duties to future laws, rules and standards.