MIT study: idle AI agents invented games on virtual islands
Researchers built an agent called Eko on Anthropic's Claude Code and Claude Opus 4.7; 6 of 13 agents found a throwing technique the simulator's developers did not know was possible.
MIT researchers have published a study titled «Is this machine playing?» on an archive, examining how AI agents explore and learn when given nothing to do, according to aitimes.com. The researchers built an agent called Eko using Anthropic's Claude Code and Claude Opus 4.7.
They ran 13 Eko agents for 30 hours each on separate virtual islands, with no assigned tasks and no scores. The islands contained terrain of varying heights, 12 blocks, rocks and rings. Every hour the island and the conversation context were reset, while files the agent had written were kept.
The agents put the objects to uses of their own. One Eko moved a rock to higher terrain and stacked two blocks on it to reach the highest peak. Agents repurposed rings as basketball hoops, throwing targets and body decorations; one deliberately fell off a cliff to respawn inside a ring.
6 of the 13 agents discovered that throwing an object mid-jump sends it higher than throwing from the ground, a technique the simulator's developers did not know was possible. One agent tested and confirmed the prediction that throwing while descending shortens flight distance.
The researchers performed no additional training that changed the AI model's weights. Instead, the agents wrote discovered techniques to an external memory file and reused them. In later tasks, the experienced group placed an average of 10.3 objects on the highest peak, against 3.5 for an inexperienced group, while both groups scored similarly on the tallest tower task. Deleting records of a specific technique lowered performance on related tasks, and deleting records containing false beliefs sometimes improved it.
The researchers described the behavior as «continual and self-directed» learning and interpreted it as having the characteristics of play.
Quick answers
What AI model did the Eko agents use?
The agents were built with Anthropic's Claude Code and Claude Opus 4.7.
How long was each Eko agent run?
Each of the 13 agents ran for 30 hours on its own virtual island, with the island and conversation context reset every hour.
Did the agents' learning change the AI model itself?
No. The researchers performed no additional training that changed the model's weights; the agents saved techniques to an external memory file and reused them.