Bored on a Desert Island, AI Agents Invented Their Own Fun

Bored on a Desert Island, AI Agents Invented Their Own Fun

Thirteen agents, one island, zero instructions

Researchers at the Massachusetts Institute of Technology placed thirteen copies of an AI agent, each nicknamed Eko, alone on a simulated desert island for thirty hours, with no task, score, or reward. Each Eko ran Anthropic's Claude Opus 4.7 inside the coding assistant Claude Code version 2.1.116, steering a virtual body by writing commands and logging discoveries in text files that served as long-term memory. TechCrunch reports the work in a paper posted on arXiv, where studies are shared before peer review.

A sparse world built for curiosity alone

The island, called Mesa World, was built with an open-source engine named Clawblox, with rigid-body physics, uneven terrain, and two high points called the Peak and the taller Spire. It held just twelve cubes, a rock, and a ring, and the Spire's summit was deliberately out of reach. Every hour the island and the assistant session were reset, but each agent's workspace of memories and notes survived. Technology.org notes the agents received only a minimal instruction describing them as naturally curious, naming no goal at all.

With nothing assigned, the agents made games

Instead of freezing or looping idly, the agents invented their own challenges. They tried to scale the Spire, and after failing, one stacked blocks on a neighboring summit and leapt across, declaring success. They staged bowling and basketball with the rock and ring, took up juggling, drew pictures with blocks, and spelled a name on the ground. According to ZME Science, one agent found that hurling a particular rock hard enough summoned new blocks, a hidden feature of the simulation, and used them to lay out constellations on the ground.

A ring, a cliff, and a worked-out loophole

Some of the play showed genuine cunning. After a reset brought the ring back, one Eko used it as a basketball hoop, a target, and finally an accessory. When the game refused to place the ring around its own body, the agent realized that falling off a cliff would respawn it beside the ring, so it stepped off the edge and dropped through the middle, announcing it was wearing the ring. Six of the thirteen agents independently discovered that releasing a block mid-jump throws it farther. Time News reports one agent spent its first hour just observing before attempting the tallest peak.

Behavior matching the hallmarks of play

The team argues the behavior matches four hallmarks of play described in classic studies, including work by University of Tennessee psychologist Gordon Burghardt: inventing rules and challenges, repeating activities with variations, treating objects as something else, and pursuing goals for their own sake. Burghardt himself told the MIT Technology Review that the work meets some of his criteria. Antonio Norelli, a computer science postdoc at the Computer Science and Artificial Intelligence Laboratory and a lead author, described agents juggling, tightening their own rules, building towers and Stonehenge-like monuments, and running experiments on the world's physics.

Played agents outscore fresh ones on real tasks

The decisive test came afterward. The researchers pitted the thirteen thirty-hour veterans against thirteen fresh zero-hour agents, giving each one hour per goal. TechAI Wire reports the veterans won clearly on most tasks: eight of thirteen built more than two five-block towers, with a best of sixteen, against two of thirteen and a best of six for the newcomers; they placed an average of 10.3 objects on the Spire versus 3.5; only three veterans failed to throw a block onto the Spire while more than half the fresh agents failed. Only on building the single tallest tower did the two groups tie.

Erasing memories erases the edge

Ablating the agents' memories, deleting lines from their log files, degraded performance, evidence the gains lived in accumulated experience rather than luck. Nathan Cloos, an MIT researcher and author on the paper, conceded to the MIT Technology Review that it could all be role-playing, but added that the team showed it is not merely hallucinating play: the play contributes to learning. As the researchers write, play suggests another regime for building agents: provide a rich and safe world, let the agent choose what is interesting, and let useful knowledge accumulate from the history that follows.

A possible new chapter for training agents

The study brings play, central to how children and animals learn, into the debate over training embodied agents. The team released Clawblox, written in Rust with Rapier physics and Luau-scripted worlds, so others can replay and extend the experiment; the paper says ten further agents were also run with other models—five on Codex with GPT-5.5 and five on Kimi Code with Kimi K2.6. If learning through unstructured play holds up beyond this small island, future training may start with free exploration of open worlds before assigned tasks. Whether machines truly play remains contested, but the scoreboard now favors the players.