MIT left 13 AI agents alone on virtual islands for 30 hours — they started to play, and it made them better
alliekmiller · x · 2026-10-24
An MIT preprint (authors include Jacob Andreas, Daniela Rus, Phillip Isola) placed 13 embodied agents powered by Claude Opus 4.7 on simulated islands with no task — only a short soul file reading "You are naturally curious. You get bored easily." Each ran for 30 hours.
- Agents invented their own games: bowling with a rock, spelling their name in blocks, self-juggling with escalating rules (higher throws, more objects)
- One agent never climbed anything, having written itself a note that climbing was "a dead end"
- Every hour the island reset and context was wiped, but memory files persisted
- Key finding: agents that had played generally beat identical agents that hadn't on subsequent building challenges
The paper asks whether this behavior satisfies classical criteria for play and whether play could become a mode of machine development.
Related event: MIT Strands 13 Claude Agents on an Island for 30 Hours, They Start to Play(2 posts)→
More from Research
- Cisco ships Agentic FAITH: one workflow for cybersecurity agent evals — aminkarbasi · 2026-10-24
- 1979 paper dismissed formal verification; 2026 forecast says it'll be as common as fuzzing — danbri · 2026-10-24
- NVIDIA and UW Protein Design team up to release RFD4-Proteina for designed biology — AllThingsApx · 2026-10-24
- T²Quant: task- and timestep-aware mixed precision for quantizing diffusion LMs — Ofek Avivi · 2026-10-24
- Berkeley & AWS: agents return faster code with Lean proofs of equivalence, 1.85x mean speedup — xiuyu_l · 2026-10-24
- Researcher: perplexity evals favor code-only models; pass@k would be far more informative — eliebakouch · 2026-10-24