Overview
CoworkOS: Scaling Agent Coworkers for Real-world Automation
CoworkOS is a multi-agent harness for long-horizon, human-in-the-loop task execution. It supports both complex one-off delivery and sustained supervision, with people able to inspect results and redirect activity throughout. Learnability, autonomy, and scalability guide the design.
Benchmarks
Performance across benchmarks.
Showcase
RSI
Recursive self-improvement.
CoworkOS turns evaluation feedback into the next training run—a working loop toward recursive self-improvement.
Five rounds. Eight benchmarks. 33.19 → 42.95.
Contributors
Core Contributors
Jiajie Jin1, *, Guanting Dong1, *, Chengjun Pan3, *, Xin Deng2, Zhuohui Sheng2, Guoqiang Wang2, Tao Gui2, †.
Contributions by Area
General Capability Optimization
Yujia Chen2, Shihong Yi2, Binze Hu2, Zhixiong Yang2, Zhichao Sun2, Rui Wang2, Han Han2, Mingyu Ma2, Junhao Huang4, Xiaotao Wang2, Guoqiang Wang2, Huicheng Jiang2.
User Experience Optimization
Yuyang Hu1, Xinyu Yang1, Wanghan Xu4, Yujia Chen2, Shihong Yi2, Guoqiang Wang2, Yajie Yang2.
RSI Optimization
Jiajie Jin1, Guanting Dong1, Zhuohui Sheng2, Xin Deng2, Zhixiong Yang2, Han Han2, Mingyu Ma2.
Harness Development and Maintenance
Jiajie Jin1, Chengjun Pan3, Yuyang Hu1, Zhuohui Sheng2, Guanting Dong1.
Discussion and Support
Zhecan Wang, Honglin Guo2, Xiaoyu Xing2, Xiaoqian Zhang6, Yulun Zhu2, Xiaoxi Li1, Jiahang Lin2, Shaofan Liu2, Shihan Dou2, Zhiheng Xi2, Qi Zhang2.
Project Mentorship
Tao Gui2, †, Yu-Gang Jiang2, See-Kiong Ng5, Yang Zhang5, Tat-Seng Chua5, Yutao Zhu1, †, Ji-Rong Wen1, Zhicheng Dou1, †.
Affiliations
1 Renmin University of China; 2 Fudan University; 3 Peking University; 4 Shanghai Jiao Tong University; 5 National University of Singapore; 6 East China Normal University.
* Equal contribution (co-first authors).† Corresponding authors.