MIT's AI agents set their own challenges on a virtual island

TechCrunch reports that MIT researchers left 13 Claude-powered Eko agents, each alone, on a simulated island for 30 hours. With no assigned task, they set challenges such as climbing a spire and using a ring as a hoop. On later tasks, the island-experienced agents generally outperformed identical agents without that experience. TechCrunch also cites a cognitive scientist who told MIT Technology Review that the paper does not establish whether the models were genuinely playing.