According to CoinMeta, the project jointly proposed by Tencent and Fudan University is dedicated to building complex, verifiable work training environments for AI and agent. A total of 16,665 training environments have been constructed, covering 40 professional fields and 60 file types. Each workspace contains an average of 42.3 files, making them more complex than simulation environments generated solely by AI. In financial tasks, agent involves analyzing stock factor returns from different regions, organizing data, and delivering it to PPT, Excel, and CSV. workforge extracts verifiable facts from the original files and automatically generates task and scoring criteria. After training, a 35B model's score on gdpval increased from 45.5 to 73.6, while another 27B model reached 82.4 on gdpval, slightly higher than qwen3.5-397b (79.7), but still lower than the 90.1 achieved by gpt-5.6 and SOL.