MIT CSAIL's 'Eko' Trial: 30 Hours of Play, Measurable Skill Gains

Agents left alone on a virtual island for 30 continuous hours outperformed untrained copies on follow-up tasks, according to an MIT CSAIL preprint posted in October to arXiv. The paper documents instances of Anthropic's Claude Opus 4.7, nicknamed "Eko," producing behaviors the authors argue satisfy four hallmarks of play.

The team, led by postdoc Antonio Norelli with coauthors Nathan Cloos and professor Phillip Isola, deployed each agent on a Roblox-style island for 30 hours — roughly a standard workweek. The agents received a "soul file" describing them as "naturally curious" and a recurring continuation prompt reading: "Please don't stop. Just continue doing whatever you want to do."

What did the agents actually do?

Each agent logged its activities in text files that served as long-term memory. They explored the island's edges, attempted to climb the highest peak, and invented goals such as throwing blocks upward.

One Eko discovered it could spawn new blocks by manipulating a larger rock, then began arranging them on the ground. It started with hearts, moved to pyramids, and eventually rendered its own name and constellations such as Cassiopeia. "Tonight I built the night sky in stone," that agent wrote in its journal.

A different Eko logged 50 consecutive juggling catches, then cancelled the record for being too easy and restarted with stricter standards.

Does the behavior meet formal definitions of play?

The authors map their observations onto a checklist developed by Gordon Burghardt, a University of Tennessee psychologist whose animal-play work the paper cites. The agents invented their own rules, repeated tasks with variations, treated objects as other objects, and pursued goals for their own sake.

Norelli treats the juggling episode as definitive. "This is playing," he said. "You give yourself a challenge and you try to achieve it, and then you make it harder."

How robust is the underlying claim?

After unstructured time on the island, agents tested on assignments such as ferrying objects to the highest peak scored higher than fresh copies. Deleting lines from their memory files erased much of that advantage.

Norelli frames this as a potential third training regime. "There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards," he said.

External reviewers urged restraint. Eunice Yiu, a Berkeley cognitive scientist, argued behavioral evidence alone cannot distinguish genuine play from imitation. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said.

Cloos conceded the alternative reading. "All of this could be role-playing," he said. "But we show that it's not just hallucinating the play. This play is also contributing to learning."

Isola attributed the behavior to cultural training rather than spontaneous motivation. "It's not like this is a system that just suddenly magically decided to play," he said. "It's an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us."

Reviewers have not yet peer-reviewed the paper. The "soul file" priming likely nudges agents toward human-like behavior, and sessions often included long stretches of inactivity that required the continuation prompt to resume.

What does this mean for autonomy roadmaps?

The authors position playful learning as a candidate third pillar alongside supervised and reinforcement approaches.

Yiu argues the field underinvests in exploration relative to exploitation. "There's a lot of work that's just on exploiting AI, making it better and faster," she said. "But in terms of building world models and exploration, I think that [they are] still a big part that deserves attention, especially if we want to generalize and prepare ourselves for new environments."

For R&D managers planning autonomy roadmaps for embodied systems, the preprint offers a hypothesis worth budgeting for replication, not yet a deployment-ready result. Whether independent groups confirm that unstructured activity compounds into measurable skill will decide if playful learning moves from preprint to procurement list.