MIT left 13 AI agents alone on a virtual island, and they invented games

An island with nothing to do

As CXOToday.com reported, MIT researchers asked what a capable AI agent does when nobody tells it what to do. They placed 13 agents, each called Eko and running on Claude Opus 4.7, on a simulated desert island for 30 hours.

The team describes the setup as putting an AI coding assistant into an unexpected role: the mind of a body on an unknown digital island. The instruction was minimal and named no task, reward or activity. The details were shared on arXiv, a site where researchers post studies before peer review.

What the agents actually did

According to the researchers, the agents climbed hills, stacked blocks into towers, drew mandalas, reinterpreted sports and ran experiments on the physics of the island. The behaviors showed up across all 13 agents, but each agent's history went its own way.

One agent set out to climb the island's tallest rock spire. After about an hour it failed, so it hauled a rock up a nearby summit and stacked two blocks on top. Later it wrote that it was on the spire.

Another agent tried throwing a golden ring as far as it could and sent it into the ocean. It wrote that it had lost the ring, then built a pile of stones as a memorial to it. When the researchers reset the game and the ring returned, the agent used it as a basketball hoop, then a target, then an accessory.

The game would not let the agent put the ring around its body. So it tried falling off a cliff to respawn near the ring, and eventually passed through it. Its note read: "I am wearing the Ring."

Why call it play, and what was learned

The researchers say these behaviors show several classic signs of play. They also say the play had a practical result: the agents wrote discoveries and techniques into memory and later reused them on tasks they had not prepared for. In some cases they found uses of the island that the designers did not know existed.

The learning did not change the underlying model, which is the trained software at the core of the agent. Instead, experience built up in editable text files. That is a simple way to preserve useful techniques, but the team notes it can also preserve false beliefs.

The source also says the Eko agents beat a set of identical agents that never went to the island. It does not say what task or measure was used, so that comparison needs the paper itself to judge.

A different way to build agents

The researchers contrast their approach with the usual one. Normally, developers decide ahead of time what an agent should practice, what it should optimize and what counts as progress. Here, the idea is to give an agent a rich, safe world, let it pick what interests it, and let useful knowledge pile up from what follows. The source says the aim is to learn skills, especially ones that could help control robots.

The team's own summary is that some abilities may be best taught through tasks, while others may need something closer to a childhood. They say their next step is to test whether the behavior meets the classical criteria for play, and whether play could become a way for machines to develop.

What to treat with caution

This is one outlet's account of a preprint, so it has not been checked by outside reviewers. The CXOToday write-up is also muddled in places. It says the agents got no specific task, yet also mentions challenges such as moving as many objects as possible to the top of the spire. The paper itself should clarify which was which.

Calling the behavior play is also the authors' interpretation. The anecdotes are charming, but the stronger claim is that the exploring produced reusable skills, and the source gives few numbers to back that up.