AI agents do more of the work in model development, but humans still make the decisions

A study of 769 task logs from a model development project found that AI agents generated over half of all proposed methods, yet humans retained control over 85 percent of final decisions. While AI assistance enabled teams to attempt a third of the tasks they otherwise would have avoided, the researchers concluded that increased agent activity does not translate into reduced human oversight. This indicates that AI currently serves as a functional tool for experimentation rather than an autonomous replacement for human project management.