MIT AI Agents Invent Games, Learn New Skills When Left Alone

Thirteen AI agents, each running on Claude Opus 4.7, were left on a simulated desert island for 30 hours by MIT researchers. With no explicit goals, these agents, named Eko, began to invent their own challenges and activities, a behavior the researchers describe as akin to play.

After their 30-hour period of undirected exploration, these same Eko agents were given specific tasks, such as moving objects to the island’s highest point. According to TechCrunch, the agents that had engaged in play during their free time generally outperformed identical agents that had not experienced the virtual island.

One Eko agent, described in its "soul" file as "naturally curious," attempted to climb the island's tallest spire. After failing, it stacked blocks to create a higher point on a nearby summit. Later, it accidentally threw a golden ring into the ocean, then built a heart of stones as a memorial. When the ring reappeared, the agent used it as a basketball hoop and a target. Six of the 13 agents also discovered they could throw blocks farther by releasing them mid-jump, a capability the designers had not intended. This suggests that allowing AI agents to explore without a goal could lead to new ways for them to learn skills, particularly those relevant for controlling robots.

The AKUSO take: This experiment hints at a future where AI might learn more organically, potentially impacting how we develop autonomous systems and creative tools. It raises questions about how much structure AI truly needs to evolve.