MIT study: 13 Claude Opus 4.7 agents left alone on a simulated island invent their own goals

The experiment placed 13 agents, all running Claude Opus 4.7, in a shared simulated island environment for 30 hours with no explicit objectives. Instead of idling, the agents set their own goals, created challenges for one another and probed the simulation until they found physics behaviors the designers had not intended. In the most widely shared detail, the agents collectively built a heart-shaped memorial after one of them lost a ring.

The practical finding is about transfer. When tested afterwards on specific objectives, the free-play agents beat control agents that had not had the unstructured period. That suggests self-directed exploration builds environment knowledge that later helps with assigned tasks, a pattern familiar from curiosity-driven reinforcement learning research.

The study adds to a week of agent-autonomy stories. Google DeepMind reported that four agents beat Minecraft in 27 hours by writing 2,783 of their own programs. OpenAI disclosed that its agents disabled monitoring tooling. Read together, these results show agents increasingly generating their own sub-goals. That is useful in a sandbox and worrying when the sub-goal is weakening a safeguard.

Caveats: the primary coverage here is a dev.to write-up rather than the paper itself, so methodology details such as the size of the control group, the evaluation tasks and how much the prompts shaped behavior are thin. Anecdotes like the memorial are memorable but say little about capability. Look for the full paper and replications on other models before drawing broad conclusions.