Fudan-linked study of 700 task logs finds AI agents did more of the work of building a model, but humans still made the decisions

A research team that includes researchers from China's Fudan University studied how humans and AI agents divided the work of building a new AI model, The Decoder reports. The team analyzed more than 700 task logs from 56 participants, along with logs from the agents they used, on its own project. That project produced Atria Dawn Preview, an agentic language model with a mixture-of-experts architecture and 744 billion parameters, designed for research and engineering tasks. The model was trained with a pipeline that ties each task to a real execution environment. It calls tools, produces intermediate results and is checked against external signals such as tests, metrics or source evidence. According to the report, the study found that agents took on a growing share of the development work while humans kept responsibility for decisions. The findings challenge some expectations about how independently agents can work. On capability, the team says Atria Dawn Preview leads on five of 16 benchmarks, including AutomationBench, CyberGym and MLE-Bench Lite, but does not hold an overall edge over competing models and trails on GDPval and SWE-Bench Pro. These benchmark results are the team's own. The study is relevant to claims about automated AI research because it offers one of the few detailed records of how much of a real model-building project agents actually handled.