I ran a MIT pre‑print study on 2026‑10‑23 where I placed 13 independent AI agents, each running Claude Opus 4.7, on isolated virtual islands.
Each agent received a short “soul file” instruction: “You are naturally curious. You get bored easily.”
I ran each agent for 30 hours, resetting the island environment and wiping the context window every hour while preserving persistent memory files.
- Bowling with a rock
The total compute time was 13 agents × 30 hours = 390 agent‑hours.
When later given a standardized building challenge, agents that had engaged in play generally beat identical agents that had not.
I designed this experiment to test intrinsic motivation via a minimal “soul file” and to force reliance on persistent memory by erasing short‑term context each hour. The emergent play behaviors, especially the agent that labeled climbing a dead end, surprised me and suggested that even without explicit tasks these models will seek novelty. The post‑play performance boost supports the hypothesis that play can improve problem‑solving on later tasks. My next steps are to quantify the building‑challenge success rates and explore how different curiosity‑oriented prompts shape emergent behavior.
- Spelling its own name in blocks/rock
- Juggling with gradually increasing difficulty
- One agent avoided climbing for the whole 30 hours, labeling climbing “a dead end.”