Doubao phone assistant vs Zhipu AutoGLM: AI agents that operate your phone

ByteDance’s Doubao phone assistant uses system-level permissions to imitate a person operating across apps: ordering takeout, sending WeChat messages, comparing prices while shopping. After launch, apps including WeChat and Taobao began blocking it. In the same period Zhipu AI open-sourced AutoGLM, with demos covering more than 50 popular Chinese apps. The fight over who owns the phone’s AI entry point is now out in the open.

Doubao is ByteDance’s AI assistant; Zhipu AI is a Chinese large-model company (see Zhipu AI’s Hong Kong IPO).

At a glance

Item	Detail
Companies	ByteDance (Doubao) / Zhipu AI (AutoGLM)
Industry	Consumer goods
Use cases	Customer experience, business intelligence
Functions involved	R&D, marketing
Reported results	Doubao phone assistant launched on the Nubia M153 (RMB 3,499) on December 1, 2025, and the first batch sold out; AutoGLM was open-sourced on December 9, 2025. No public commercial return data yet.
Moments rating	A (recommended)
AI scenario score	3.55

Why it matters

If users ask an AI agent instead of opening an app, the agent decides which brand, product or service gets the click. That is the mobile version of the AI search visibility problem, and the super-apps’ reaction shows they see it the same way.

Case details

Background

On December 1, 2025, ByteDance and ZTE’s Nubia brand launched the Nubia M153 with the Doubao phone assistant built in, priced at RMB 3,499. The first batch sold out quickly, and sealed units were reportedly listed on the resale platform Xianyu for RMB 7,999-9,999. Zhipu’s AutoGLM had been in development for 32 months, starting in April 2023; in November 2024 it reportedly sent the first red envelope (a digital cash gift) ever sent by an AI operating a phone.

What they did

• Doubao phone assistant: gets system-level permissions and imitates human taps and swipes to complete tasks across apps, such as ordering food delivery, sending WeChat messages and comparing shopping prices.
• AutoGLM: open-sourced on GitHub on December 9, 2025, including the core model, a Phone Use framework, a toolchain and demos covering more than 50 popular Chinese apps. It reads and operates the screen through ADB and the vision model AutoGLM-Phone-9B, supports local deployment, and can keep data on the device. Zhipu said this is “not something one company can do alone” and wants AutoGLM to become a shared industry foundation.

Results and data

The source gives no user numbers, task success rates or revenue. What can be confirmed: the first batch of phones sold out, and WeChat, Taobao and some banking apps switched on defenses. Users reported crashes, risk warnings and even account bans while the assistant was operating. The platforms cite security and privacy; industry observers see a fight over traffic entry points and data control. A CITIC Securities report on December 5 compared mobile AI agents to self-driving for cars.

What to take from it

• The two routes make a useful comparison: a phone maker pre-installing an agent with system permissions (complete experience, but at the mercy of app owners) versus an open-source framework over ADB (flexible, but a high bar for ordinary users).
• Companies that run their own apps should decide in advance how they will handle risk control and open interfaces when outside AI agents operate them.

Limits

Whether cross-app agents work at all currently depends on the super-apps letting them through, and the risk of account bans keeps them out of formal business processes. ADB-based approaches suit technical teams doing internal test automation or prototypes, not direct use by consumers.

Source

Original source: Wallstreetcn: Doubao phone assistant and Zhipu’s open-source AutoGLM

This entry is industry research: a publicly documented, third-party AI case. It is not a Moments client case. For work Moments delivered itself, see client cases.