China is building its AI regulatory stack from both ends at once.

Last week gave us two separate China AI governance moves that, read together, tell a more interesting story than either does alone.

Let me try to unpack both properly.

The State Council legislative plan

The 2026 Legislative Work Plan (gov.cn/zhengce/content/…), approved by the CPC Central Committee, confirms "comprehensive AI legislation" as a priority. That phrase has been floating around Chinese policy circles for a couple of years now, so the confirmation matters less than the specifics of what it's supposed to cover.

The listed scope: data, computing power (算力), algorithms, IP, network security, and supply chain security.

Most coverage stopped at "AI legislation confirmed." But the 算力 inclusion deserves its own sentence. Computing power getting explicit legislative mention in a State Council plan is new. China has been building out its national compute infrastructure aggressively, and it's also acutely aware, post-2022 chip export controls, that its compute supply chain runs through jurisdictions that have demonstrated willingness to cut access. Legislating around 算力 isn't just about AI governance. It's about treating compute as a strategic asset that needs legal protection and, probably, legal mechanisms for state prioritisation. That's a different kind of AI law than anything the EU, UK or US is building.

Supply chain security appearing in the same list is the same logic applied more broadly. No other jurisdiction's AI legislative framework explicitly names supply chain security as a target. The EU AI Act doesn't. The US Executive Orders touched on it but through a national security lens, not a legislative AI-specific one. China embedding this into a comprehensive AI law would be genuinely novel.

Then there's the form question, which the plan doesn't answer and which matters enormously. China's existing AI rules, the Algorithm Recommendation Measures (2022), the Deep Synthesis Measures (2022), the Generative AI Measures (2023) are at the lower tier of the Chinese regulatory hierarchy, being CAC-issued administrative rules. They bind, but their enforcement power and their ability to override other frameworks is limited.

A State Council legislative plan points toward something higher: either a 行政法规 (State Council administrative regulation) or, more ambitiously, a NPC standing law. The difference is significant. A State Council regulation sits above departmental rules but below NPC law. An NPC law would signal a level of political priority that would reverberate across every sector. The plan doesn't specify which track so it remains TBC.

And the plan also flags revision to the Measures for the Administration of Internet Information Services, which has been CAC's main enforcement lever for internet and AI services. If the comprehensive AI law and the revised IIS measures are being drafted in parallel, the relationship between them, which governs what, which takes precedence in a conflict, becomes a real drafting challenge. Or possibly a deliberate ambiguity that preserves CAC's flexibility.

The AI agent guidelines

On 8 May, three agencies (the CAC, the NDRC, and the MIIT) jointly released Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents (cac.gov.cn/2026-05/08/c…)

The three-agency combination is doing more analytical work than the content itself. CAC is the internet regulator. NDRC is China's economic planning body, the one that issues five-year plans and designates industrial priorities. MIIT covers manufacturing and telecoms. These three don't usually sign things together.

When China uses a multi-agency joint release, it can signal one of two things: genuine inter-agency alignment on something that crosses jurisdictional boundaries, or a negotiated compromise that gave each agency enough to sign off. The NDRC's presence specifically suggests the latter framing is closer. NDRC doesn't regulate AI. Its role is economic development planning. Its signature on an AI agent document signals that this is being positioned as an industrial development initiative, not primarily a governance or safety one. The "AI plus" branding in the guidelines makes that explicit. China ran "Internet Plus" (互联网+) from 2015 as a demand-side push to integrate internet capabilities into traditional industries. The structural logic here is identical: identify target sectors, designate them as priority application areas, use regulatory recognition to drive investment and adoption.

The guidelines identify 19 specific application scenarios. That number is very Chinese policy planning. I read them as more like “designated sectors” rather than clarificatory examples. Scientific research and public services are mentioned explicitly. That list will matter for which industries get preferential treatment in licensing, funding, and standards development.

The guidelines define intelligent agents as autonomous systems with perception, memory, decision-making, and execution capabilities. That's a broader definition than most Western frameworks use. "Execution" in particular is doing a lot of heavylifting here - i.e. it extends the definition beyond systems that reason or recommend to systems that act. Under a loose reading, a lot of existing software starts to look like an "intelligent agent" under this definition. Before a comprehensive AI law exists to provide hierarchical context, that definitional scope creates real compliance uncertainty for developers and deployers.

Which surfaces the tension I find most interesting about last week's two moves together.

• China is operationalising AI agent governance before the framework law that should sit above it exists. The agent guidelines are live now. The comprehensive legislation is a 2026 priority that hasn't been drafted yet. When that law arrives, what happens to the agent guidelines? Do they get elevated into the legal hierarchy, superseded, or left as soft guidance floating beneath a law that may use different definitions, different jurisdictional carve-outs, different enforcement mechanisms?

• Most jurisdictions, including the EU, tried to set the framework law first and let sector-specific rules follow. China is running the process in reverse, prioritising market development and operational guidance now, and sorting out legal architecture later. It’s faster but it creates layering problems that someone, probably the courts or CAC, will have to resolve under pressure.