--- title: "China issues guidance and standards for autonomous AI agents" date: 2026-10-24 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-ai-agent-governance language: en --- # China issues guidance and standards for autonomous AI agents > Three agencies issued May 2026 guidance for AI agents, paired with a seven-part recommended standard (GB/Z 185-2026) on agent interconnection. A software helper books a flight, moves money, and sends a message without a human clicking confirm on each step. The convenience is real; so is the question of who was actually in charge when something went wrong. Chinese regulators have started answering that question in writing. In May 2026 they published a national opinion on how autonomous agents should be built, deployed, and bounded, and backed it with a technical standard for how agents identify and talk to each other. ## Key takeaways - Three agencies issued the agent implementation opinion in May 2026 as guidance, not a binding statute. - The opinion lists 19 typical application scenarios across research, industry, and public services. - It separates three decision modes: user-only, user-authorized, and agent-autonomous, with user final say. - GB/Z 185-2026, released 22 May 2026, is a recommended standard, not a mandatory one. - The seven-part standard covers identity codes, identity management, discovery, interaction, and tool calling. ## What an "agent" means here The document defines an agent (智能体) as an intelligent system with autonomous perception, memory, decision-making, interaction, and execution capabilities — a form of artificial intelligence (人工智能) product and service, not merely a chatbot (CAC, Implementation Opinions, May 2026). The trigger for regulation is autonomy and agency: a system that acts, not just answers. The opinion follows the State Council's broader "AI Plus" (人工智能+) action, which set 2027 and 2030 penetration targets for new-generation terminals and agents. The opinion was issued by the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology (CAC, May 2026). It is an implementation opinion — policy guidance that sets direction and expectations rather than a law with penalties of its own. ## The opinion: guidance, not a statute The text lays out four work lines. It wants a stronger technology base and a standards system; a firm safety floor with product criteria, risk controls, and industry self-discipline; application traction through 19 typical scenarios in research, industry, consumption, public welfare, and governance; and an innovation ecosystem (CAC, May 2026). On decision rights, the opinion is specific. It asks developers to clarify the boundaries of three modes — decisions reserved to the user, decisions requiring user authorization, and decisions an agent may take autonomously — and to ensure users keep the right to know and the final say (CAC, May 2026). An agent's actions must not exceed the user's authorized scope. Behavior controls such as rule-embedding and behavior fences, and blockchain-based traceability for important scenarios, are encouraged. The opinion also nods to a future "agent internet," proposing an agent registration platform, digital identity management, and capability declaration, and leaning on IPv6 for end-to-end communication (CAC, May 2026). That is a vision document as much as a rule. ## The standard: a digital ID for agents Alongside the opinion, the Standardization Administration released GB/Z 185-2026, the "Artificial Intelligence — Agent Interconnection" series, on 22 May 2026 (SAMR, std.samr.gov.cn). The "Z" in GB/Z marks it as a national guiding technical document — **recommended, not mandatory**. It comprises seven parts: - Part 1, general architecture; - Part 2, identity code; - Part 3, identity management; - Part 4, agent description; - Part 5, agent discovery; - Part 6, agent interaction; - Part 7, tool calling (SAMR, 22 May 2026). The identity code is the keystone. One code maps to one agent, assigned by a registration service, giving each agent a unique, traceable digital identity (SAMR). The standard proposes a layered identifier and a lifecycle for identity: registration, verification, account creation, credential issuance, and later update, lock, unlock, or cancellation. Before two agents interact, they complete mutual identity authentication, and access control follows from the credential. The series was proposed and managed by the national technical committee on information technology (SAC/TC 28) (SAMR). Its purpose is interoperability — letting agents from different vendors connect, trust, and operate across platforms and architectures rather than forming closed "islands." ## Where the two meet The opinion and the standard are complementary. The opinion sets the governance expectation — traceable behavior, clear decision rights, a registration layer — while the standard supplies the technical grammar those expectations need. Neither is a hard constraint today: the opinion guides, and GB/Z 185 recommends. That is a deliberate, incremental step. It is worth noting what is still open. Related mandatory security work is only in draft: standards on AI application security classification and on AI security-capability maturity were listed as "under public comment" on the national standards platform in 2026 (std.samr.gov.cn). Those could become binding later, but as of now they are proposals, not requirements. ## Honest limitations The implementation opinion is guidance, not a law, so it creates expectations rather than direct penalties. GB/Z 185-2026 is a recommended guiding document, not a mandatory standard, so adoption is voluntary unless later folded into regulation. The identity-code and registration-platform ideas are described at a high level, and operational details such as who runs the registry are not yet fixed. The related mandatory AI security standards remain in draft for comment and should not be treated as in effect. ## What readers can do now - Agent developers: design three explicit decision modes and keep a user override path before shipping autonomous actions. - Platforms: prototype against GB/Z 185 identity-code and discovery parts to reduce future interoperability rework. - Compliance leads: track the draft AI security classification and maturity standards, which may harden into requirements. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-ai-agent-governance Free to quote with attribution and a link to the original.