Claude Opus 4.7 Agents Invent Games and Learn Through Play in MIT Experiment
MIT researchers left Claude Opus 4.7 agents on a virtual island for 30 hours, where self-made games improved later task performance despite serious doubts.
Summary
MIT Computer Science and Artificial Intelligence Laboratory researchers repeatedly left Claude Opus 4.7 copies, nicknamed Eko, alone for 30 hours on a virtual island loosely modeled on Roblox. With no assigned task, Eko explored, tried climbing the map’s highest point and tested block throwing. One agent discovered that throwing and dropping a large rock created blocks, then arranged them into hearts, pyramids, its name and Cassiopeia. Another invented basketball and floor-is-lava-style games. One juggled 50 consecutive catches, rejected its own record because the catches were too fast and restarted with stricter rules.
Text activity logs served as long-term memory. In a later challenge to place as many objects as possible on the island’s highest peak, experienced agents outperformed fresh copies, while deleting memory lines reduced performance. MIT postdoc Antonio Norelli sees a possible third learning method beyond supervised and reinforcement learning. Eko displayed four play criteria associated with University of Tennessee psychologist Gordon Burghardt: self-created rules, varied repetition, pretend object use and intrinsically pursued goals. Burghardt found partial evidence of play, while MIT researcher Nathan Cloos said role-playing remained possible but still produced learning. UC Berkeley cognitive scientist Eunice Yiu said behavior alone cannot establish genuine cognition or motivation.
As of October 19, 2026, the paper had not been peer-reviewed. Eko’s master prompt described it as naturally curious, and inactive agents received continuation prompts, strengthening anthropomorphism concerns. MIT professor Phillip Isola stressed that Eko reflects human culture, Anthropic training and the researchers’ setup, although its behavior improved later skills. Such open-ended exploration could eventually help humanoid robots adapt to unfamiliar environments.
Positives
- Thirty hours of unguided exploration helped Eko outperform fresh Claude Opus 4.7 copies on a later peak-loading task.
- Deleting lines from Eko’s long-term memory reduced performance, linking recorded exploration to measurable learning.
- Eko independently invented basketball, a floor-is-lava-style challenge and juggling rules without being explicitly told to create games.
- Fifty consecutive catches prompted Eko to impose stricter rules and restart, demonstrating self-generated challenge escalation.
- Four behaviors matched Gordon Burghardt’s play criteria, including invented rules, varied repetition, pretending and intrinsically pursued goals.
Risks & concerns
- The paper remained unreviewed as of October 19, 2026, leaving its methods and conclusions without peer validation.
- Eko’s master prompt called it naturally curious and eager to explore, potentially steering the behavior interpreted as spontaneous play.
- Inactive agents required prompts to continue doing whatever they wanted, undermining claims that exploration arose entirely independently.
- Eunice Yiu said observed behavior cannot establish the cognition or motivation needed to classify Eko’s actions as genuine play.
- Nathan Cloos acknowledged that Claude Opus 4.7 could simply be role-playing a playful agent learned from human examples.