On October 24, the General Office of China's Ministry of Industry and Information Technology issued a "Notice on Strengthening Network Access Management for Agent-Enabled Terminal Devices," document number MIIT Office Information Management [2026] No. 42. It was only made public on November 11. The notice is short — four tasks — but two of them rewrite foundational rules: for the first time, network access testing reaches into the software layer, and for the first time, model updates require prior approval.

Testing shifts from the device to the software

First, some context on what a network access license is. Telecom terminal devices sold in China must obtain a network access license from MIIT, under the Telecommunications Regulations and the Administrative Measures for Network Access of Telecommunications Equipment. Until now, the test covered the hardware itself — whether RF performance meets spec, whether it passes electromagnetic compatibility, whether the specific absorption rate stays within limits, whether it interoperates properly with the network. All hardware work.

The new notice pushes that mechanism a step further. It states that phones, tablets, wearables and other agent-enabled terminals with preinstalled agents must undergo agent adaptation verification during network access testing, in line with the relevant standards. They must also undergo fraud-risk security assessment and upfront anti-fraud capability evaluation, as required by law.

What gets tested at the factory gate is no longer just whether the board works, but whether the agent inside it does.

The notice defines an agent as "an intelligent system with autonomous perception, memory, decision-making, interaction and execution capabilities," and an agent-enabled terminal as "a telecommunications terminal device such as a phone, tablet or wearable with a preinstalled agent."

The scope hinges on one word: preinstalled

"Preinstalled" is the critical word in this document.

Agent apps that users download themselves from an app store fall outside the definition — those are covered by the fourth task, handled at the distribution stage. Network access testing applies only to what the manufacturer ships by default.

The split is deliberate. Preinstalled means default capability: the user can use it from the moment the device is switched on, without having to decide whether to install it or judge whether it is safe. If a default capability goes wrong, the manufacturer is responsible, so it gets stopped at the licensing gate. If the user chose it, responsibility is shared through that choice, and the oversight sits at the distribution end.

The fourth task is that distribution half: app distribution platforms should push app developers to use electronic signatures to actively declare whether they consent to being invoked by agents. Without a declared consent, an agent may not forcibly invoke or access an app. Distribution pages must publish adaptation verification lists, and platforms may not offer users agent downloads for incompatible device models. Platforms must also audit listed agents, requiring those that have not completed adaptation verification to rectify within a deadline or be delisted.

One rule covers what leaves the factory, another covers what goes on the shelf. The stretch in between — how an agent behaves once the user has installed it — is not addressed this time.

Updates need approval, and this is the hardest clause

The wording of the third task is heavy.

Terminal manufacturers that update agents on live networks via over-the-air pushes or similar methods must submit an upgrade plan to MIIT, specifying timing, scale, user groups, use cases, service content, business model, network and data security measures, and anti-telecom-fraud measures. After the upgrade, they must submit a report, and only after MIIT organizes expert assessment and solicits comments can the service be formally offered to users.

Phone system updates used to run on the manufacturer's own schedule: finish the build, push it out, and users wake up to a new version. That no longer works.

The biggest casualty is iteration speed. Large-model products now typically ship monthly or even weekly, adding capabilities incrementally and fixing issues incrementally. That cadence cannot survive a "submit plan — wait for assessment — solicit comments" pipeline.

The phrase "may be formally offered to users only after assessment and comment solicitation" also leaves an opening: before formal release, can a company run a staged rollout? Can it push to beta testers? The document does not say. That is the first question implementation will run into.

What counts as a major change goes undefined

One line in the second task reads: terminal manufacturers shall ensure consistency of the agent-enabled terminal before and after it obtains the network access license. If an agent-enabled terminal undergoes a major change, the manufacturer must file with MIIT and conduct network access testing as required.

"Consistency" is a standard clause in licensing regimes, meant to prevent companies from submitting one product for testing and selling another. But agents are peculiar in that they are built to change.

Does swapping the underlying base model count as a major change? Does going from 7 billion parameters to 30 billion? Does adding a persistent voice wake-word that keeps the device always listening? Does turning a question-and-answer assistant into an executor that can operate other apps on the user's behalf?

The document draws no line. And that line directly determines how many times a manufacturer has to run network access testing — once per model swap versus one filing a year are two very different cost structures.

The identity code plugs into a standard set two months ago

Another line in the second task is easy to skip: agents used on terminals should be assigned a unique identity code with reference to the national standards series on agent interconnection.

That series is GB/Z 185.1 through 185.7 — seven parts covering overall architecture, identity identification, trusted management, capability description, intelligent discovery, multi-modal interaction and tool invocation. The standards have already been promoted in the Yangtze River Delta and elsewhere, with sessions specifically covering identity code encoding rules and the full process by which an agent applies for a code, is verified, obtains credentials and registers on a platform.

The word choice is worth noting. In the same notice, adaptation verification is to be conducted "in line with the relevant standards," while the identity code is to be assigned "with reference to" the national standards series. GB/Z documents are national standardization guiding technical documents — neither mandatory nor recommended, and carrying no binding force of their own. Writing "with reference to" means the technical content of the standards has been pulled into network access testing, while their legal status remains unchanged.

The practical effect of that combination: the standards remain voluntary, but you cannot pass testing without them.

The supporting measures include one more line: accelerate construction of a trusted identity system and management platform for agents. As of the notice's issuance at the end of October, that platform was still being built. In other words, the identity code requirement now has a basis in policy — but nowhere yet to actually issue the codes.