China sorts every agent decision into three tiers of authority
Human-only, user-approval-required, agent-autonomous. Before deployment, each of an agent's decisions must be assigned to one. It is the first national regime to treat autonomy as a permission level rather than a property of the model.
China's Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents took effect on 15 July 2026 and are widely described as the first national framework treating AI agents as a distinct regulated category. The operative requirement: before deployment, each of an agent's decisions must be sorted into one of three tiers — human-only, user-approval-required, or agent-autonomous.
That mechanism is a more consequential idea than the rest of the document. Most AI regulation so far has regulated the model: what it was trained on, what it can output, what must be disclosed. This regulates the action, and it does so by forcing a decision that most deployments currently leave implicit.
Anyone who has shipped an agent knows the implicit version. The boundary between what it does on its own and what it asks about is usually a prompt instruction and a handful of conditionals, written by whoever built it, documented nowhere, and adjusted when it becomes annoying. A rule requiring that boundary to be enumerated in advance turns an engineering convenience into a filed artifact.
It is also, unusually for AI regulation, straightforwardly implementable. A tier assignment is a table. Compare that with obligations to characterise training data or demonstrate the absence of a capability, which impose real cost precisely because nobody agrees what compliance looks like.
The obvious caution is that a framework's text and its enforcement are different things, and the same document also serves industrial policy aims that have nothing to do with safety. But the three-tier construct is portable, and Western regimes currently regulating models rather than actions should expect to meet it again. California's bill slate is still working the older seam.
Inside AI Policy — AI regulation coverage, August 2026 → · EveAI Core — AI Regulation in 2026: What Just Changed — and What It Means for Every AI Team →