The pledge gives China's largest agent builders a shared baseline on security, user control and interoperability before binding rules arrive, and sets expectations that regulators can later point to.
Twenty Chinese organizations, including Tencent, Huawei, Alibaba, Ant Group and Douyin, have signed a voluntary convention setting out how AI agents should be secured, governed and connected. The Agent Development Self-Discipline Convention (《智能体发展自律公约》) was released on November 4 at a forum on agent interoperability during the World Internet Conference’s Wuzhen Summit, with the China Academy of Information and Communications Technology (CAICT) among its drafters, The Paper reported. CAICT is a research institute under China’s Ministry of Industry and Information Technology.
The convention contains eight self-discipline measures. They cover trust systems for agents, shared infrastructure for connecting them, interoperability, security, governance, user rights, adoption across industries, and wider access to agent technology. The signatories pledge to apply the measures in their own products first and to accept supervision from industry bodies and the public, according to MyDrivers.
The document has no enforcement mechanism beyond those pledges, and neither report gave a timetable for meeting its commitments. Its weight comes from the list of names: between them, the signatories run many of the agent products Chinese consumers and businesses use, from chat assistants to phone operating systems.
The signatories span platforms, phone makers and security firms
The convention was drafted under the organization of the China Federation of Internet Societies (中国网络社会组织联合会), and CAICT co-drafted and released it with the other sponsors, MyDrivers reported, citing CAICT’s official WeChat account. Besides CAICT, Tencent, Huawei, Alibaba, Ant Group and Douyin, the named signatories include phone makers Honor, OPPO and Xiaomi; Baidu and Du Xiaoman; AsiaInfo Technologies; China Mobile Internet; Changan Communication; security companies Sangfor, Qi An Xin and 360; and Shenyan Intelligence. MyDrivers named 18 of the 20 sponsors in its text; the full list appeared only in an image.
The mix matters for interoperability. Phone makers are building agents that act across apps, while platforms such as Tencent and Alibaba control the apps those agents need to enter. The convention’s third measure calls for interoperability within legal and security limits, with respect for intellectual property, and through open collaboration and equal consultation. It lists interface adaptation, task delegation, tiered authorization, defined behavioral boundaries and the allocation of responsibility as areas to work on, The Paper reported.
MyDrivers described the convention as a concrete step in carrying out the industry self-discipline task set by the Implementation Opinions on the Standardized Application and Innovative Development of AI Agents (《智能体规范应用与创新发展实施意见》), the national guidance on agents issued in May by the Cyberspace Administration of China, the National Development and Reform Commission and the Ministry of Industry and Information Technology.
Security and governance clauses carry the most specific language
The fourth measure, on security, asks signatories to build forward-looking risk assessment, runtime monitoring, behavior controls and "abnormal circuit-breaking" systems, to ensure models, plugins and skills come from reliable sources, and to share risk information and respond to incidents jointly. It names the risks to guard against: agents overstepping their permissions, leaking private data or losing behavioral control because of flaws in the AI itself or outside attacks.
The fifth measure, on governance, calls for technology that embeds rules into agents and tests their compliance, so that agent behavior follows laws and public order. Tencent, Huawei, Alibaba and the other signatories commit to jointly resisting the use of AI agents for algorithmic exploitation and black- and gray-market activity. The Paper’s headline highlighted that clause.
The user-rights section commits signatories to protect users’ right to know, right to choose and final control over important decisions, and to prevent misuse of personal data, excessive pandering to users and cognitive manipulation, with extra protection for minors and older people.
Product changes and new signatories will show whether it bites
The convention invites other companies in the agent ecosystem to join, so the first signal is whether more developers, such as AI model startups and large internet platforms not yet named, add their names.
The more telling evidence will be in products. Visible circuit breakers, permission prompts that let users approve an agent’s important actions, or published agent registration and discovery standards from CAICT would show the signatories turning the text into practice. If regulators later cite the convention when drafting binding rules for agents, its eight measures will have served as the first draft.