AutoGLM

AutoGLM is a new member of Zhipu AI's ChatGLM family, focused on autonomously controlling digital devices through graphical user interfaces (GUIs). Serving as a practical foundation agent framework, it can understand and operate device interfaces such as web browsers and Android phones.

Key Features

• Phone Agent: Understands phone screen content through multimodal perception and uses ADB to control the device, automatically executing users' natural language instructions. Supports 50+ mainstream Chinese apps spanning social, e-commerce, travel, video, and other scenarios.
• Web browser control: Achieves a 55.2% success rate on the WebArena-Lite benchmark and 96.2% on OpenTable evaluation tasks—roughly a 200% performance improvement over GPT-4o.
• AutoGLM Rumination: China's first desktop agent with deep research capability, able to break problems into steps and autonomously open browser tabs to search, find, record, summarize, and analyze information. Information retrieval efficiency is 40% higher than ordinary AI, with a 92% task execution accuracy rate.

Use Cases

On mobile, users can give AutoGLM voice commands like "search Meituan for hotpot restaurants nearby" or "send a WeChat message," and the system automatically opens the relevant app and completes the whole workflow. On desktop, AutoGLM Rumination handles complex tasks requiring cross-platform information integration—academic research, market research, information synthesis—breaking through traditional AI's information silos.

Unique Advantages

AutoGLM achieves a 36.2% success rate on the AndroidLab benchmark for Android device control and 89.7% on common Chinese app tasks. The version released on October 25, 2024 is widely considered the world's first AI agent with genuine Phone Use capability. The system has a built-in sensitive-operation confirmation mechanism and supports human takeover for login or CAPTCHA scenarios to ensure safety.