China now requires human sign-off before an AI agent can act — sorted into three tiers

China's Cyberspace Administration, National Development and Reform Commission, and Ministry of Industry and Information Technology have jointly issued Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents — the first national policy framework written specifically for AI agents rather than generative AI in general. Multiple policy trackers report the measures reaching compliance effect on July 15, 2026, the same week as China's separate rules on anthropomorphic companion AI.

The framework defines an agent as a system capable of autonomous perception, memory, decision-making, interaction and execution — a deliberately narrower category than a chatbot, built around the fact that an agent can take real-world action rather than only produce text. Its central mechanism is a three-tier authorization structure. Some decisions remain user-exclusive, requiring direct human control regardless of the agent's confidence. Others are authorized decisions, delegated to the agent through explicit user permission — a standing grant, not a one-time click. The remainder are autonomous decisions the agent may take within its defined scope, though the framework states users 'retain the right to know and the ultimate decision-making authority' even over actions they never individually approved.

Filing, testing, recall — for the sectors where mistakes matter most

Agents operating in 'sensitive sectors and key industries' face heavier obligations: formal filing with regulators, compliance testing, and product-recall provisions if an agent is later found unsafe. The sectors named explicitly are healthcare, transportation, media, and public security — the categories where an autonomous action, taken wrong, is hardest to undo. An organization deploying an agent in one of those sectors without a documented decision-authorization policy is, per the framework, already out of compliance.

Users retain the right to know and the ultimate decision-making authority — even over the decisions the agent is permitted to make on its own.

A different category from the companion-AI rules

It would be easy to fold this into China's separate companion-AI rules, since both reach compliance effect in the same week. They are not the same instrument. The companion-AI rules regulate emotionally engaged interaction — AI designed to be treated as someone. This framework regulates the opposite category: AI built to work for you, sorted by how much authority it holds when it does. Read together, China has now drawn two distinct regulatory lines through consumer-facing AI in a single week — one for AI that bonds, one for AI that acts — where most jurisdictions have drawn none.

What is not established

The dating is messier than the compliance-deadline headlines suggest. One source traces the document's issuance to May 8, 2026, with no separate effective date specified in the text itself; multiple trackers instead report July 15 as when compliance obligations bite. The precise issuance-to-effect timeline could not be independently confirmed against China's official gazette, and no source located names a specific penalty schedule or enforcement body empowered to act on a violation yet. The examples given for each authorization tier — financial transactions, government approvals, medical decisions, judicial functions — are illustrative, not an exhaustive list; how a regulator classifies a genuinely ambiguous agent action in practice is untested.