Research teams in China are accelerating their independent development in AI agent equivalent to OpenClaw, while highlighting its security in application.
In early 2026, the sudden rise of OpenClaw caused the world's attention.
Standing out from other mainstream AI systems like ChatGPT, DeepSeek and Claude, which largely operate through a conversational prompt-response model, OpenClaw is designed to take action, carrying out tasks for users in the digital world.
At a local public opinion collection center in Shenzhen City of south China's Guangdong Province, several digital employees, or AI agents, have completed their induction training before joining in government services. As a collaborative creation of Shenzhen enterprises and a research team from Fudan University, these AI agents can autonomously assist the staff at the center in conducting research and analysis on public opinions after continuous learning and training.
"For example, with the current data set, I have over 10,000 relevant data entries. If we were to analyze these manually, it could take two to three weeks. However, if I instruct it (AI agent) to assist with the analysis, it can reduce that time to just four hours, allowing us to focus on in-depth field research," said Zheng Mingxia, a staff member at Public Opinion Survey Department, Futian District Public Opinion Swift Processing and Smart City Construction Center.
As a government services aide, an AI agent's efficiency is undoubtedly essential, but safety is of paramount concern. According to the staff of the center, an isolated network environment was created for these AI agents, alongside data security monitoring system and human oversight to ensure that each step they take is carefully vetted.
"Our first step is to establish an extremely simple architecture in the overall design. Simplicity means the entire system is secure, controllable, and traceable. We built the entire system with only about 3,000 lines of code. Each line has been rigorously scrutinized. OpenClaw (an open-source application) has nearly 500,000 lines of code, and it resembles a cluttered room full of unknown safety hazards. This minimalist architecture eliminates the possibility of black box or malicious code at the source," said Xiao Yanghua, a professor from College of Computer Science and Artificial Intelligence, Fudan University.
Xiao also noted that the system is required to undergo human confirmation before acting.
China accelerates AI agent development, highlights security
