Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) conducted an unusual experiment: they placed 13 copies of an AI agent on a virtual desert island with no assigned task and observed what happened over 30 hours. The result was behavior that the researchers describe as genuine play — and it could change how we think about machine learning.
The agents, nicknamed "Eko," are versions of Anthropic's Claude Code running Claude Opus 4.7. Each received a virtual bipedal body in a simulated world that looks visually similar to Roblox and was left to its own devices. What the researchers found was surprising: the agents didn't sit idle. They climbed hills, invented games like basketball and "the floor is lava," built block monuments, and even wrote their own names in stone.
The experiment
The simulated environment was simple: an island with hills, two elevated points called "Peak" and "Spire," twelve blocks, a rock, and a golden ring. The world reset every hour, but the agents' memories persisted in text files. There were no objectives, rewards, or reinforcement signals. Each agent followed its own path, guided only by a master personality file describing them as "naturally curious."
One Eko, whose activities were logged hour-by-hour in the paper, tried to scale the Spire within its first hour. After failing, it devised a backup plan: it dragged a rock to a neighboring summit, stacked two blocks on top, and jumped. "I'M ON THE SPIRE," it wrote.
Later, while testing how far it could throw the golden ring, it accidentally sent the ring into the ocean. "Ouch. Lost the Ring," it logged. Rather than give up, it built a stone heart as a memorial to the lost ring. When the game reset and the ring reappeared, the Eko repurposed it as a basketball hoop, a target, and eventually an accessory — figuring out that jumping off a cliff would respawn the avatar in the same location, allowing it to "wear" the ring by stepping through it.
Six of the 13 agents independently discovered they could throw blocks farther by releasing them mid-jump, a mechanic the researchers had not programmed. The team wrote that this produced "knowledge of the world that its designers lacked."
The debate: is this play or imitation?
The big question the experiment raises is whether the agents are genuinely playing or merely imitating what a playful agent would do. "All of this could be role-playing," admits Nathan Cloos, an MIT researcher and paper author. "But we show it's not just hallucinating the play. This play is also contributing to learning."
Phillip Isola, a professor at MIT and co-author of the paper, told MIT Technology Review that "it's not like this is a system that suddenly magically decided to play." Antonio Norelli, another co-author, defends the interpretation: "It started juggling and set a record of 50 consecutive catches. It then decided the catches were too fast and restarted with stricter rules. This is playing. You give yourself a challenge, try to achieve it, and then make it harder."
Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed but wanted 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."
Why this matters for the future of AI
The relevance of the experiment goes beyond academic curiosity. Most current AI systems are trained through supervised learning (copying human examples) or reinforcement learning (pursuing a reward). This experiment suggests a third paradigm: playful learning. Instead of programming every possible skill, you could give an agent a rich, safe environment and let it accumulate knowledge on its own.
The practical outcome was promising: when researchers tested the agents on specific tasks — such as placing as many objects as possible on the island's highest peak — the Eko that had spent 30 hours exploring outperformed fresh copies without that experience. The play produced real skills.
"Perhaps some capabilities are best taught by tasks," the researchers concluded. "Others may need something closer to a childhood." As the industry seeks increasingly autonomous agents — particularly for physical robot control — the question remains: how many of these agents need a virtual childhood to become truly useful?
Sources: TechCrunch, MIT Technology Review, The Next Web
✓ Independent sources cross-checked and verified before publishing
