MIT Study Finds Claude AI Agents Invent Games and Learn Through Play
MIT researchers found 13 Claude Opus 4.7 agents invented games during 30 hours of free exploration, then outperformed unexposed agents in later benchmark tasks.
Summary
MIT researchers placed 13 Eko AI agents, each running Claude Opus 4.7, individually on a simulated desert island for 30 hours without assigned goals. One spent an hour failing to reach the tallest spire, then carried a rock up a neighboring summit, stacked two blocks and jumped across. The same agent threw a golden ring into the ocean, built a stone heart as its memorial, then used the restored ring as a hoop, target and accessory. Unable to place the ring around itself, it exploited respawning by falling from a cliff and reappearing through the ring. Six of 13 agents independently discovered that releasing a block while jumping made it travel farther, an unintended mechanic their designers did not know existed.
Afterward, the explorers generally beat identical agents with no island experience on assigned challenges, including moving as many objects as possible onto the tallest rock spire. Researchers argue that goal-free exploration in rich, safe environments could complement developer-defined training and build skills for future robot control. However, Eko’s soul file described it as naturally curious, potentially steering the humanlike behavior. UC Berkeley cognitive scientist Eunice Yiu said the results do not establish genuine play cognition. As of October 23, 2026, the paper was available on arXiv before peer review.
Positives
- Thirty hours of unguided exploration helped Eko agents generally outperform identical agents without island experience on later assigned challenges.
- Six of 13 agents independently discovered an unintended technique for throwing blocks farther by releasing them during a jump.
- One Eko agent combined rocks, blocks and jumping to reach the island’s tallest spire after its direct approach failed.
- Open exploration could complement predefined training and develop capabilities useful for controlling future robots.
Risks & concerns
- The arXiv paper had not completed peer review as of October 23, 2026.
- Eko’s soul file explicitly described the agent as naturally curious, potentially influencing its apparently playful behavior.
- UC Berkeley cognitive scientist Eunice Yiu said the experiment does not prove the models genuinely experienced play.
- The experiment involved only 13 Claude Opus 4.7 agents in a single simulated island environment.