The People's Daily proposed on November 16 that agile governance should safeguard the development of AI agents, calling for a classification and grading governance framework, exploration of sandbox-based and trigger-based regulation, and filing, testing and recall management for sensitive domains.

For low-risk domains, the newspaper said compliance self-testing, information reporting, distribution platform management and industry self-discipline should play the leading role. It also clarified the boundaries and permissions of three decision modes: decisions limited to the user themselves, decisions requiring user authorization, and autonomous agent decisions.

Analysts note the regulatory focus is formally shifting from what models say to what agents do. The Economic Daily quoted Wang Jian, founder of Alibaba Cloud, as saying open source does not mean there are no boundaries, while 360's Ye Jian observed that when a model errs it may say something wrong, but when an agent errs it may do something wrong.