China's Zhipu AI launches open-source phone agent

Zhipu AI has released AutoGLM, an open-source AI agent model that can operate smartphones by interpreting on-screen content and simulating user actions.

AutoGLM can handle multi-step tasks such as placing food delivery orders and booking flights by performing taps, swipes, and text input.

It currently works with over 50 widely used Chinese apps, including WeChat, Taobao, Douyin, and Meituan.

The open-source release includes trained models, a phone-use framework, toolchain, Android adaptation layers, runnable demos, and documentation.

Zhipu AI said the project supports both local and cloud deployment to maintain user control over data and privacy.

🔗 Source: TechNode

🧠 Food for thought

Implications, context, and why it matters.

AutoGLM licensing will shape real-world deployment

On its site, the team lists a Creative Commons Attribution-ShareAlike 4.0 International license for the website code 1. It allows commercial use and requires share-alike. The materials do not specify a license for the AutoGLM code or model. That gap could constrain original equipment manufacturers (OEMs) and app builders that need closed-source integration.
Docs say it’s “intended for research and learning only, prohibiting illegal use or system interference” 2. That stance could complicate production deployments by hardware makers, such as smartphone vendors that ship devices, unless clarified.
Absent clear commercial terms, phone makers and developers face legal risk when building AutoGLM into consumer products. That uncertainty keeps the impact narrow beyond academic work.

Reliability platform for mobile AI agents as UIs keep changing

In AndroidWorld, a research benchmark suite that programmatically generates mobile tasks, the M3A agent reached 30.6% task success 3. Agents struggle with perception errors and UI element interaction failures.
Product teams that build on AutoGLM could use a monitoring service that tracks task success across top apps and flags UI changes as apps update weekly.
An MVP should target apps AutoGLM already supports. Use AndroidWorld’s dynamic task generation to run continuous tests and surface failure modes before they hit production 3.