On November 6, China's Ministry of Industry and Information Technology (MIIT) published a notice that brings phones, tablets and wearables with pre-installed AI agents under its network-access regime. It requires compatibility testing, a unique ID for each agent, reporting of over-the-air (OTA) upgrades, and bars agents from forcibly invoking an app unless the app's developer has declared consent. For device makers and agent developers, the competition is shifting from model capability towards compatibility, compliance and ecosystem coordination.

What happened

The notice from MIIT's General Office (document No. 42 of 2026, dated October 24) defines an "agent terminal" as a phone, tablet, wearable or other telecom device with an AI agent pre-installed. The notice cites China's laws on cybersecurity, data security, personal information protection and telecom fraud, together with the Telecommunications Regulations and the existing rules on network access for telecom equipment, and extends that network-access management to the agents on board.

The notice sets out four tasks:

  • Compatibility: agent developers must list, in the product's settings page, the device models and operating systems the agent has been verified on, and report this verification to app stores before listing.
  • Network access: devices with pre-installed agents must undergo agent compatibility verification during network-access testing, plus a fraud-risk assessment and an anti-fraud capability evaluation. Agents on the device should carry a unique ID with reference to China's national standards on agent interconnection. Makers must report major changes to the ministry and retest as required.
  • Upgrades: device makers that upgrade agents over the air must file an upgrade plan covering timing, scale, user groups and business scenarios; afterwards they file a report, and the service can formally launch only after an expert review organised by the ministry and a round of consultation.
  • Distribution: app stores should have app developers declare, by electronic signature, whether they consent to being invoked by agents. Without that declaration, agents may not forcibly invoke the app. App stores must not offer agents for incompatible device models, and must order fixes or delist agents that have not completed verification.

Background

The notice reads more clearly against two threads from the past year.

The first is standards. In July 2026 China released GB/Z 185-2026, a set of seven guidance-level national standards (the GB/Z prefix marks guidance documents) on AI agent interconnection, covering overall architecture, agent identity codes, identity management, capability description, cross-domain discovery, collaborative interaction and tool invocation. More than 2,000 agent identity codes for key industries were issued at the launch event, according to CCTV News, the news service of China's state broadcaster. The new notice asks agents on devices to carry a unique ID with reference to these standards.

The second is a dispute. On December 5, 2025, Doubao Phone Assistant, an agent that can operate a phone on the user's behalf, announced limits on its own capabilities: restricting automated actions used to collect points and rewards in apps, further restricting its operation of banking and payment apps, and pausing AI use in some ranked competitive games. Users had earlier reported being blocked from logging in to WeChat and other third-party apps when using its phone-operation feature, and one point of contention was whether a user's authorisation can substitute for the third-party platform's own, according to The Paper (Pengpai News).

Hongshugu Insights

The agent turns from a feature into a regulated object. Compatibility checks become part of network-access testing, major changes must be reported, and OTA upgrades need a plan beforehand, a report afterwards and an expert review before launch. For device makers used to shipping new features continuously over the air, agent release planning, compatibility testing and regulatory filings are likely to become fixed steps in product development, and the ability to coordinate versions and compliance will shape how fast new features reach users.

Consent moves from platform stand-offs to written declarations. Last year's dispute left open whether user authorisation can replace a platform's authorisation. The notice's answer is a declaration by the app developer, signed electronically. How much an agent can do for its user will therefore depend largely on how many apps choose to consent, and negotiations between agent makers and the largest apps may become the next arena of ecosystem competition.

For technology suppliers, a higher bar and a new division of labour. Published compatibility lists, unique IDs and fraud-risk assessments raise the bar for agent developers to get onto devices, while turning compatibility testing, identity management and security assessment into distinct specialist services. For smaller developers, the ability to work with device makers and app stores may matter as much as the model itself. Large enterprises planning to deploy agents on employee devices may also find verification status and agent IDs a useful item in supplier evaluation.