MIT researchers let 13 Claude Opus 4.7-based agents independently explore a simulated island for 30 hours before assigning specific tasks. Reportedly, experienced agents generally outperformed otherwise identical newcomers and discovered uses of the simulation’s physics that designers had not anticipated. However, the base prompt described the agents as curious, and this is a single preprint study. Playlike behavior does not establish emotions or humanlike cognition, and transfer to real robot learning remains to be tested.