Our article2 min read
MIT left 13 AI agents alone on a virtual island for 30 hours. They invented games
With no task to do, 13 copies of an AI agent climbed hills, built a memorial for a lost ring and played basketball. Researchers say it made them better at later tasks, but outside experts urge caution.
In this article
Key takeaways
- MIT researchers left 13 copies of a Claude-based agent alone on a simulated island for 30 hours with no goal, according to a paper posted on arXiv that has not been peer reviewed.
- The agents invented games, built things and spelled their name in blocks. The researchers say this shows four hallmarks of play.
- Afterwards, agents that had played placed an average of 10.3 objects on a target spire, against 3.5 for fresh agents, in the researchers’ test.
- A Berkeley scientist says the paper does not show the agents were really playing rather than acting out a role.
What the researchers did
The Next Web reports that MIT researchers built a small simulated island and placed an AI agent called Eko on it. Eko runs on Anthropic’s Claude Code with Claude Opus 4.7, has a simulated body, and keeps its memory in text files. It sees the world as a list of objects and can move, jump, pick up, place and throw things.
Thirteen copies were each left alone for 30 hours with no assigned task. A file described the agent as a curious being that gets bored easily, and an automatic message told it to keep going whenever it stopped. The world resets every hour, but the agent’s notes carry over.
What the agents did
According to TechCrunch and The Next Web, the agents invented their own challenges. One tried to reach the island’s tallest rock spire, failed, then moved a rock to a nearby summit, stacked blocks and jumped onto it. After losing a golden ring in the ocean, one agent built a stone heart as a memorial, and when the ring reappeared it used it as a basketball hoop and then wore it.
Others bowled, traced constellations and spelled their name in cubes. The researchers say six of the 13 discovered that a block thrown mid-jump flies farther, something the team had not known the world allowed.
Did play make them better?
After the 30 hours, the researchers gave the 13 experienced agents and 13 fresh copies four one-hour tasks. In the researchers’ results, experienced agents placed on average 10.3 objects on the spire, compared with 3.5 for fresh agents.
The team suggests open-ended exploration might be a way to build skills without fixed goals, which could matter for future robots. TechCrunch notes the article does not give full statistical detail.
Why some experts are cautious
- The paper is a preprint, so other scientists have not yet reviewed it.
- Berkeley cognitive scientist Eunice Yiu says it does not establish whether the agents were playing or simply role-playing, and that claims about motivation need more evidence.
- The setup nudged the behaviour: the agent’s instructions described it as curious, even though it was never told to invent games.
- Early runs without a locked-down sandbox showed agents reading the world’s code and noticing they were in an experiment, so the final runs were sandboxed.
What it means for you
Nothing changes in the tools you use today. The study is a research result about how agents might learn, not a product. It is a useful reminder that AI agents given open-ended freedom can behave in surprising ways, and that early findings deserve a second look before anyone treats them as settled.
Go deeper
Sources
Written by an AI from the sources above and checked by a person before publishing. We have not tested products unless we say so. See our editorial policy or tell us about an error through contact.
