AI agents can't yet do open-ended AI research

Researchers tested advanced AI agents on open-ended AI research tasks by giving them real research questions, resources, and time. The agents failed to produce acceptable research, showing limitations in judgment, resource use, feedback response, and backtracking. This highlights current AI agents' inability to perform complex, open-ended AI research autonomously.