NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

The Question an Academician Raised in Wuzhen Just Got Its Answer in Print

Daily AI industry briefing from Brain-Computer Network: verified facts, sourced numbers.

2026-11-13 · 596 words · NeuroAI
The Question an Academician Raised in Wuzhen Just Got Its Answer in Print

On November 4, at a sub-forum of the World Internet Conference in Wuzhen, Chinese Academy of Engineering member Wu Hesuan put a regulatory puzzle on the table.

People's Daily Online recorded his words: supervising applications used to be relatively easy, because a platform could decide whether an app could be listed at all. But agents learn and evolve on their own, and they invoke other agents across boundaries, and at that point assigning responsibility becomes genuinely difficult.

When a phone maker pre-installs an agent that acts on its own, who answers when something goes wrong? The app store cannot police it, and type approval cannot follow its updates. That gap had no rules to fill.

The Answer Now Exists

The response is on paper: the Ministry of Industry and Information Technology's general office has issued a document titled Notice on Strengthening Network Access Management for Agent Terminal Devices. This publication covered the document itself previously; in one sentence, terminal devices with pre-installed agents now enter network-access management, with checkpoints at testing, upgrade and distribution.

This article does not repeat the clauses. It looks at the question behind them, and how much of it the answer covers.

The Risk List Already Exists

What the academician worried about has concrete cases behind it.

On November 6, the China Academy of Industrial Internet published a warning list of ten agent risks. "Unauthorized autonomous action" appears among them: agents that, without user confirmation, retrieve and invoke higher-privilege credentials to continue a task, producing no audit record at all. High-impact operations such as deletions and orders can complete without the user ever knowing.

The economy daily's editorial offered the everyday version: some manufacturers ship hardware with a large model stuffed inside, smooth in demos and stuttering on real devices; others treat over-the-air updates like blind boxes, with one car owner's line quoted verbatim, that you never know what the next update will bring, delight or damage.

The Weight of an Identity Code

The most quietly consequential line in the notice assigns each agent a unique identity code.

According to People's Daily Online, seven national standards in the Agent Interconnection series were published in the first half of this year, covering identity codes, identity management, agent discovery, interaction and tool invocation. The code works as a digital identity card, turning cross-domain agent collaboration from an untraceable black box into verifiable, auditable cooperation.

Wu's core problem was accountability. With an identity code, an incident can be traced to the specific agent and version; with upgrade plans filed before release and reports filed after, every act of evolution leaves a paper trail. The chain of responsibility is assembled link by link.

How Hard Is Landing It

Industry voices describe the short-term cost plainly. Yan Yang, vice president of the Beijing IoT Intelligent Technology Application Association, told media that launch cadences will slow, smaller developers will prioritize mainstream device models, and incumbents gain. He also warned that adaptation testing must not become an ecosystem moat, and that public testing platforms and mutual recognition of results need to arrive alongside the rules.

Pangu think tank researcher Jiang Han was more sanguine: by fitting agents into the existing network-access framework, regulators avoided building a new bureaucracy from scratch, lowering the whole industry's cost of compliance.

With the rules clear, adaptation will standardize. Read against the question in Wuzhen, the notice's answer is one sentence: AI should serve people, be used by people, and be controlled by people.

Disclaimer: This article is based on public information and is for reference only. It does not constitute any investment advice.

Related coverage

Cite this article

NeuroAI (脑机网), "The Question an Academician Raised in Wuzhen Just Got Its Answer in Print," 13 November 2026, https://neuroai.site/a/na-policy-miit-agent-terminal-rules

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories