MIT's Eko Agents Performed Better After Exploring a Simulated Island

MIT researchers let 13 AI agents called Eko explore a simulated island for 30 hours without assigning them a task. Each agent ran Claude Opus 4.7. In a later challenge to place objects on the island’s tallest rock spire, the agents that had explored generally did better than identical agents that had not.

Open-ended exploration could help agents acquire useful skills, including skills that may eventually apply to robots. The approach differs from training that sets an agent’s task and measure of success in advance. Eko’s instructions described it as “naturally curious,” but did not tell it to invent games.

One Eko agent first tried to reach the spire and failed. It then carried a rock to a nearby summit, stacked two blocks on it and jumped, reaching the spire. Later, while testing how far it could throw a golden ring, the agent sent it into the ocean. It responded by arranging stones into a heart as a memorial to the lost ring.

The agents also found a way to throw blocks farther: Six of the 13 released a block while jumping, a maneuver the researchers had not designed. The discovery was one example of the agents learning how objects behaved through exploration.

UC Berkeley cognitive scientist Eunice Yiu said the behavior did not establish whether the models were genuinely playing. “It could be the case that the models are genuinely playing, but we don’t know based on the paper,” Yiu said. She said more evidence about the thinking behind the behavior was needed.