MIT runs 13 Claude agents unattended on a simulated island

On October 23, 2026, a DEV Community report on AI described an experiment in which MIT researchers left 13 AI agents running on Claude Opus 4.7 alone on a simulated island for 30 hours. No tasks were assigned to the agents. According to the report, the study was published on a research-sharing platform. Rather than simply waiting for the time to pass, the agents invented their own challenges, discovered unintended physics mechanics, and later outperformed control agents on specific test objectives. The report also noted that one of the agents initially tried to climb the highest rock on the island, but the attempt failed; the summary cuts off at that point and does not give the rest of the account. The available information stops at this description of the study. The material does not explain the full experimental procedure, the specific test objectives, or how the control group was set up.