On July 15, 2026, China became the first country to carve out a dedicated regulatory category for AI agents — the autonomous software systems that do not merely answer questions but perceive, plan, and act on a user's behalf. The centerpiece is the Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents, jointly issued by the Cyberspace Administration of China (CAC), the National Development and Reform Commission (NDRC), and the Ministry of Industry and Information Technology (MIIT). It arrives alongside a second, genuinely binding measure that took effect the same day, and together they mark the moment agentic AI stopped being governed as an extension of generative AI and became a named object of Chinese law.
The distinction between the two instruments matters, and it is where much of the coverage has overstated the case. The Implementation Opinions are, by their form, a guiding policy instrument — "Opinions" in the Chinese administrative lexicon set direction, define categories, and signal the trajectory of future enforcement rather than functioning as a self-executing statute with its own penalty schedule. What gives them teeth is that they impose concrete obligations on agents deployed in sensitive sectors — healthcare, transportation, media, and public safety — including mandatory filing, compliance testing, and product-recall provisions. The separately issued Interim Measures for the Administration of Anthropomorphic AI Interaction Services, which also took effect July 15, is the strictly binding regulation of the pair, governing emotional chatbots and virtual companions through CAC algorithm filing, security assessments, and addiction-prevention rules.
The three-tier framework
At the core of the agent Opinions is a decision-authorization structure that sorts an agent's actions by consequence. Developers must classify decisions into three tiers: those the user makes, those that require the user's explicit authorization, and those the agent may take on its own. The governing principle, repeatedly stressed in the text, is that the human user retains final decision-making power and that an agent may not act beyond its authorized scope.
In practice, analysts have mapped that structure onto escalating levels of oversight. Routine, low-stakes actions — scheduling, data retrieval, basic customer-service replies — can run autonomously. Decisions with material impact, such as pricing changes or content-moderation calls that affect user access, require a human in the loop to approve. High-stakes actions carrying legal, financial, or safety consequences must escalate to a human decision-maker rather than being executed by the agent at all. Companies are expected to document which tier each deployed agent operates in and maintain audit trails, with the tightest scrutiny — pre-deployment review and ongoing monitoring — reserved for the highest-risk deployments.
The Opinions define AI agents expansively, as intelligent systems capable of autonomous perception, memory, decision-making, interaction, and execution. They identify nineteen priority application scenarios spanning scientific research, manufacturing, healthcare, and consumer services, and they pair a "safety first" posture with an unmistakable industrial-policy goal: 70 percent adoption of intelligent agents in smart terminals by 2027.
Sarah Zhao, a partner at Rimon Law who tracks Chinese AI regulation, wrote in the firm's China AI Law Brief that the Opinions "signal that China is beginning to regulate AI agents as a distinct governance category rather than simply treating them as an extension of generative AI." She added that "future regulation will increasingly focus on the deployment and governance of autonomous AI systems in higher-risk sectors" — a reading that frames July 15 as the opening move of a longer regulatory build-out rather than a finished rulebook.
Why It Matters
China is first, but it is not moving in a vacuum. The framework lands as the EU AI Act enters its August enforcement phase, extending general-purpose and high-risk obligations that increasingly implicate agentic systems, and as US states legislate piecemeal — Illinois mandating third-party safety audits for covered AI, California and New York policing anthropomorphic companion services. The contrast is one of philosophy. China's agent rules are prescriptive, spelling out authorization tiers, filing steps, and sector triggers, where the EU AI Act is broader in scope but leaves more to harmonized standards, and the US has no comparable federal instrument at all.
The structural problem each jurisdiction is chasing is identical: organizations are deploying agents with real decision-making power into oversight systems that were never built to track them. By naming the category in law and forcing companies to prove where each agent's authority ends, China has produced the first template others can copy, adapt, or define themselves against. Extraterritorial reach sharpens the stakes — agent deployments that touch Chinese users, Chinese data, or Chinese market operations can trigger filing obligations even without a physical presence in the country, pulling multinational compliance teams into scope whether or not they run infrastructure on the mainland.
What to Watch
Three questions will determine how much the template actually binds. First, enforcement detail: the Opinions lean on existing data-security and AI-governance penalty frameworks, with specific fines and mechanisms left to forthcoming implementation guidance — watch for the first recalls, cease-operation orders, or filing rejections in the sensitive sectors. Second, whether foreign platforms serving Chinese users treat the three-tier classification as a genuine compliance obligation or a paper exercise. Third, whether Brussels, Washington, or US state legislatures borrow China's core move — regulating agents as a distinct legal category with graduated authority limits — or continue folding them into broader AI statutes. China got there first; the open question is whether "first" becomes "standard."
"The Opinions signal that China is beginning to regulate AI agents as a distinct governance category rather than simply treating them as an extension of generative AI."— Sarah Zhao, Partner, Rimon Law