MIT left 13 AI agents on a virtual island, and they started inventing games

TechCrunch has reported an experiment by MIT researchers in which 13 AI agents were left on an isolated virtual island with nothing to do. The question was simple: when agents have no assigned task, what do they do?

They did not idle. According to the report, the agents invented games and eventually staged something like mourning for a lost ring. With no human-set objective, they generated their own activities and rituals.

The more consequential detail came afterward. TechCrunch says the agents ultimately performed better at their jobs, suggesting that the social and playful behavior during unstructured time fed back into their task performance.

That offers a different lens on agent evaluation. Most benchmarks measure task completion, tool-use accuracy or cost — in other words, how efficiently an agent behaves when it is told what to do. This experiment looks at what happens when nothing is asked, a period that standard evaluations rarely capture.

It also points to a practical feature of multi-agent systems: when several models share an environment over time, social dynamics, imitation and self-invented rules emerge on their own. Those dynamics can support cooperation, but they can also produce drift that is hard to predict.

Caution is warranted. The public summary does not say which models were used, how long the experiment ran, or how improved job performance was measured. It is better treated as an observation worth following than as a result that generalizes on its own.

The next things to watch are whether the full method and data are released, whether other teams can reproduce the findings, and whether idle-time behavior becomes a standard dimension in agent safety and product design. If it does, the way agents are evaluated will need a substantial addition.