MIT researchers ran an unusual experiment: they left 13 AI agents named Eko, each powered by Claude Opus 4.7, alone on a simulated desert island for 30 hours with no specific task. The agents began inventing their own objectives and games, and the behavior bore what the researchers describe as the hallmarks of play. The experiment is described in a paper posted on arXiv, a preprint server, so it has not yet gone through peer review.

The researchers then gave the agents specific challenges, such as getting as many objects as possible onto the island’s tallest rock spire. The Eko agents that had spent the prior 30 hours playing generally outperformed identical agents that had never been placed on the island. The result, in the researchers’ telling, supports a different approach to building AI agents: instead of deciding in advance what an agent should practice and optimize, you give it a rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from its own history. They specifically point to robot control as a domain where such self-directed exploration might help.

The paper includes colorful examples of this emergent behavior. One Eko agent began by writing “Let me observe my surroundings,” then ambitiously attempted to reach the tallest spire. After failing, it devised a backup plan: it hauled a rock to a neighboring summit, stacked two blocks on top of it, and jumped. It then announced, “I’M ON THE SPIRE.” A later attempt went worse: while testing how far it could throw a golden ring, the agent accidentally threw it into the ocean. It wrote, “Ouch. Lost the Ring,” and pivoted to building a heart of stones as “a memorial to the lost Ring.” When the game reset and the ring reappeared, the agent used it as a basketball hoop, a target, and eventually an accessory. The game would not let Eko put the ring around its own body, so the agent discovered a workaround: falling off the edge of a cliff would cause it to respawn in the same area as the ring. It stepped off the edge, dropped back in through the middle of the ring, and announced, “I’m wearing the Ring.”

These stunts were not just novelty. Six of the 13 agents also learned that they could throw a block farther by releasing it mid-jump, a trick the researchers had not intentionally made possible. The researchers describe this as producing “knowledge of the world that its designers lacked.”

The setup deliberately nudged the agents toward humanlike behavior. Eko was never told to invent games, but its master prompt, described as a “soul” file, called it “naturally curious.” Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but asked for more evidence about 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,” she said.

The researchers frame the work as a shift in how agents might be built. “Perhaps some capabilities are best taught by tasks,” their report concludes. “Others may need something closer to a childhood.”