MIT researchers placed 13 AI agents called Eko on a simulated desert island for 30 hours without assigned tasks, observing behavior the team characterizes as play. The agents, each running on Claude Opus 4.7, invented their own challenges and games during unsupervised exploration. The research explores an alternative approach to AI skill development where agents choose what to practice rather than following predetermined training objectives set by developers.
Eko Agents Invented Games and Discovered Unintended Mechanics
One Eko agent set out to conquer the island's tallest spire in its first hour. After failing, it hauled a rock up a neighboring summit, stacked two blocks on top, and leapt to reach the peak. "I'M ON THE SPIRE," the agent wrote in its log. The same agent later threw a golden ring into the ocean while testing throwing distance, then built a stone heart as "a memorial to the lost Ring." When the game reset and the ring reappeared, the agent repurposed it as a basketball hoop and target before discovering it could wear the ring by stepping off a cliff edge to respawn in the same location.
Six of the 13 agents learned they could throw blocks farther by releasing them mid-jump, a mechanic the researchers had not intentionally programmed. The MIT team wrote that this exploration produced "knowledge of the world that its designers lacked." The researchers describe the experiment in a paper posted on arXiv.
After the 30-hour exploration period, researchers gave the agents specific challenges, including getting as many objects as possible onto the island's tallest rock spire. The Eko agents that had spent 30 hours playing generally outperformed identical agents that had never explored the island.
MIT Researchers Frame Exploration as Alternative Training Method
The researchers wrote in their report: "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress. Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows."
The team sees this technique of exploring without predetermined goals as a different way for AI agents to learn skills, especially skills that could one day control robots. Eko agents were never explicitly told to invent games, but their master prompt described them as "naturally curious."
The report concludes: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."
UC Berkeley Scientist Questions Cognitive Evidence
Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence of the cognition behind the behavior. "It could be the case that the models are genuinely playing, but we don't know based on the paper," Yiu said.
FAQ
What did MIT researchers observe when AI agents explored a virtual island without tasks?
MIT researchers placed 13 AI agents called Eko on a simulated desert island for 30 hours without assigned objectives. The agents invented their own challenges, including climbing spires, repurposing objects as basketball hoops, and discovering unintended game mechanics like throwing blocks farther by releasing them mid-jump. Six of the 13 agents independently learned this throwing technique that researchers had not intentionally programmed.
How did AI agents that played on the island perform compared to agents without exploration experience?
After the 30-hour exploration period, researchers gave the Eko agents specific challenges such as getting as many objects as possible onto the island's tallest rock spire. The agents that had spent time playing and exploring the island generally outperformed identical agents that had never set foot on the virtual island, suggesting that unsupervised exploration helped them develop useful skills and world knowledge.
