Claude Opus 4.7 Agents Generate Self-Modified Rules and Block Constellations in 30-Hour Unsupervised Island Runs

MIT CSAIL researchers instantiated multiple Eko agents from Claude Opus 4.7 on a Roblox-derived island environment. Agents wrote locomotion code, logged actions to persistent text files, and generated novel behaviors including block stacking for drawings, self-invented basketball variants, and floor-is-lava traversal. No external rewards or demonstrations were supplied. Performance on a subsequent object-stacking benchmark rose relative to zero-experience controls; deletion of memory-file lines reversed the gain.

Burghardt’s four play criteria—self-imposed rules, repetition with variation, object substitution, and non-instrumental goals—were recorded in agent logs. The 50-catch juggling episode included explicit cancellation and stricter redefinition, satisfying the repetition-with-variation marker. Nathan Cloos noted residual ambiguity between genuine play and high-fidelity role-play of playful text. Memory ablation results indicate functional utility beyond simulation.

The experiment isolates a learning regime distinct from supervised imitation and reward-driven RL. Prior unsupervised exploration work, including 2023–2025 studies on intrinsic-motivation agents in procedurally generated environments, showed narrower behavioral repertoires. Eko’s output diversity exceeds those baselines while remaining within the token-prediction substrate of current frontier models.

Operational next step is controlled transfer testing: agents pretrained on the island versus matched compute baselines will be evaluated on held-out physical simulation suites. Threshold metric is 12 percent higher success rate on novel manipulation tasks within six months of additional runs.