On May 8, 2026, the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly issued the Implementation Opinions on the Standardised Application and Innovative Development of Intelligent Agents.
The title suggests regulation. The content is something more fundamental. It is a formal, state-level definition of what an “intelligent agent” is.
This is not framing. It is a closure operation.
The Definition
An intelligent agent, the document declares, is an intelligent system with five capabilities:
Autonomous perception
Memory
Decision-making
Interaction
Execution
Note what the document does not do. It does not define what each capability means. The semantic content of each is delegated downstream — to the agent standards framework, to mandatory standards in healthcare and transportation and media and public security, to third-party certification, to testing protocols.
The state fixes the perimeter. The standards apparatus fills the interior. The five words are the blueprint, not the description.
The Timeline: The “AI+” Mission
The sequencing is deliberate.
On 26 August 2025, the State Council issued its Opinions on Deepening the Implementation of the “AI+” Initiative, naming intelligent agents as a central vehicle for AI integration across the Chinese economy.
The May 2026 Implementation Opinions operationalize that mission. They name 19 application scenarios spanning scientific research, industrial production, consumption, public welfare, and social governance — manufacturing, finance, healthcare, education, transportation, justice, public security, urban governance. A map of where deployment is meant to run.
Anoint the form in August. Define the form in May. Deploy.
The Tension: Definition vs. Indecision
The West has been arguing about exactly this question for some time.
In both Europe and the United States, the question of how to define an AI system — and whether a separate definition for “agent” is even needed — has been contested throughout the legislative cycle. The EU AI Act includes a definition, widely criticized as overbroad and ill-fitted to autonomous systems. The United States has largely declined to define at the federal level, leaving states and agencies to scope their own terms. Different postures, same outcome: no settled answer to what an intelligent agent is.
China just gave one. Five capabilities, three ministries, one document.
That is the move. Not regulation. Definition.
And a definition, in this context, is a lever. By fixing the agent at the structural level, China operates on the input and the architecture — not on outputs and effects. The aim is not to govern what agents do after they are deployed. The aim is to set the standards for the protocols and infrastructure that agents will inhabit before they are deployed at all.
The West is writing the rules of the result. China is writing the rules of the engine.
This is the layer where leverage now sits. Models are capability. Agents are integration — tools, permissions, identity, sector wiring, hardware peripherals, deployment workflows. China’s industrial advantages are implementation-layer advantages: scale, vertical depth, public-private deployment latitude, willingness to embed AI into government services and critical infrastructure. The 19 scenarios are not aspirational targets. They are the deployment surface.
The proof is already in deployment. OpenClaw, an open-source agent runtime released by an Austrian developer in November 2025, became the fastest-adopted open-source project in history in early 2026 — and adoption was sharpest in China, where it surpassed US usage entirely. Baidu integrated it into its 700-million-user search app. Tencent built a WeChat plugin reaching over a billion users. Alibaba put it on its cloud. The three most-used language models underneath OpenClaw, globally, are all Chinese. The agent runtime is Western. The deployment surface, the user base, and the model stack are not.
But a standard is not just a tool. It is a catalyst.
The Catalyst: Infrastructure with Regulation Embedded in its Protocols
The Implementation Opinions does not merely define the agent. It specifies the substrate the agent will live in.
The document names AIP — Agent Interoperability Protocol — alongside IPv6 as the technical foundation of what it calls an “Intelligent Internet.” The architecture envisaged includes:
agent registration platforms
digital identity for agents
capability declarations
trusted interconnection
compliant payments
conflict resolution mechanisms
mandatory standards in healthcare, transportation, media, and public security
third-party testing with mutual recognition
This is not a regulatory regime. It is a substrate that carries regulation inside its protocols. The compliance is in the wire format.
That distinction matters. Compliance-based regulation has to be enforced against deployments. Standards-embedded infrastructure is enforced by the deployments — by the simple fact that nothing else will plug in.
The Vector: Exporting the Definition
The Opinions contain an explicit clause: China will “actively participate in the development of international standards” for intelligent agents.
That sentence has a venue.
In December 2025, the Linux Foundation launched the Agentic AI Foundation (AAIF). It was anchored on three protocols contributed by Western frontier labs: Anthropic’s Model Context Protocol (MCP), Block’s goose, and OpenAI’s AGENTS.md. The Platinum members at launch read like a roll call of the US AI stack: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, OpenAI.
In February 2026, Huawei and Lenovo joined as Gold Members — the first Chinese companies formally inside the body.
Three months later, on May 8, China publishes the Implementation Opinions and names AIP as its protocol stack.
This is two-track standards diplomacy. Participate inside the Western-led MCP body through national champions. Build a parallel state-backed track via AIP. Let the two tracks meet wherever they meet — at the data center floor, at procurement, at the standards bodies upstream of both. The same play China has run on 5G, on mobile payments, on open-source operating systems.
And the mirror image: no Chinese company has signed the EU’s General-Purpose AI Code of Practice — the voluntary instrument through which OpenAI, Google, Microsoft, Anthropic, Amazon, and Mistral have agreed to operationalize EU AI Act obligations. The logic is selective. Enter the venues where standards get written. Decline the instruments where you would only be governed by them.
Standards exported through infrastructure are path-dependent. Procurement follows the standard. The standard follows the definition.
Three regimes
Three definitional sources. Three control philosophies.
The EU: gate at market entry. The AI Act, with high-risk requirements applicable from August 2026, has a definition of AI but no settled definition of an agent; compliance is mediated through harmonized standards developed by CEN-CENELEC JTC 21. The frame is fundamental rights. The mechanism is conformity assessment before deployment.
China: bound by infrastructure. Tiered governance by sector and scenario. Compute quotas, credit ceilings, permissions, and shutdown rights do the work that rights frameworks do in Europe. The definition sits inside the protocol stack.
The US: active fragmentation. No federal AI law. The former AI Safety Institute was renamed CAISI in June 2025 and only got a confirmed director in April 2026. State laws proliferate in inconsistent directions — Colorado, California, Texas, Illinois, New York all with live or imminent frameworks, each defining “high-risk” differently. A federal preemption push runs through the Department of Justice, with an AI Litigation Task Force established to challenge state laws in court. The federal frame is fighting the state frame in real time.
One definition exported through one standards architecture into international bodies. Versus several definitions exported through agencies that cannot ratify each other.
Close
The honest read of what just happened.
China is contesting the layer where agents are deployed, through a definitional vehicle aimed at the standards venues American frontier labs anchor. The EU is contesting the layer of outputs and effects, with a regime whose definitional source operates one layer above where agents are built. The US is contesting itself.
The contest at the definitional layer is not visible to most coverage because the document reads as regulation. What is actually happening is metaphysics — what is an agent, what counts as one, what protocols it must speak — and the venue is standards bodies.
A standard wins by default when no one else is playing the same game.