AutoGLM

AutoGLM is an autonomous task Agent introduced within the ChatGLM model family. Rather than answering a question and stopping there, it is designed to operate graphical user interfaces on a user's behalf, taking a stated goal and working through the steps needed to reach it insIDE web browsers and Android APPlications. The project frames this as a step toward more general assistants that can ACT in messy, real-world digital environments.

What it does

Given a natural-language instruction, AutoGLM interprets the goal, plans a sequence of interface actions, and then executes them. Examples described by the project include drafting and scheduling an emAIl, looking up a highly rated nearby cafe and starting walking directions, adding specific products to a shopping cart, booking a restaurant table, filing an issue in a project tracker, or summarizing travel tips found on a social platform. It handles both Chinese and English instructions, including mixed-language requests.

Key capabilities

• Web browsing control: navigates sites, fills forms, filters listings, and completes multi-step browsing tasks.
• Android device control: performs app actions such as composing mail, searching maps, or manAGIng a shopping cart.
• Task planning and execution: breaks a goal into ordered steps and cARRies them out in sequence.
• Environment-driven learning: improves decision-making by interacting with environments rather than relying only on static training data.
• Separated planning and acting: uses an intermediate interface layer so planning can favor flexibility while execution favors accuracy.
• Progressive training: an online, curriculum-style reInforcement approach lets the agent keep refining its behavior.

Who it is for

AutoGLM is most relevant to researchers and developers studying GUI agents, computer-use models, and reinforcement learning for decision-making. Product teams exploring automation of repetitive browser or mobile workflows may also find the dEMOnstrations useful as a reference. Because the published material is research-oriented, General consumers should treat it as an experimental system rather than a polished everyday assistant.

How to get started

The project publishes documentation and demonstrations through its official project page. Those interested can review the described capabilities and example tasks there. Availability of a live, publicly usable agent may vary and could be limited to research previews or specific environments, so check the current project materials before assuming general access.

Pricing and access

No reliable public pricing information is available for AutoGLM; any commercial terms, usage limits, or paid tiers may change and should be confirmed directly with the publisher. Access conditions, supported platforms, and feature scope can also vary over time. The agent executes actions on real accounts and devices, so users should review permissions and confirm sensitive steps manually.