Orphan page — no incoming links
title: “Agent Play Learning”
aliases: [“playful learning for agents”, “Eko experiment”]
category: “ai”
created_at: “2026-10-23”
updated_at: “2026-10-25”
sources: [“2026-10-23-ai-agents-started-playing-mit”, “2026-10-23-mit-eko-play-arxiv”]
Definition
Research line where embodied LLM agents generate self-directed activities resembling play, storing discoveries in persistent memory that later improves task performance without explicit reward engineering.
Key Points
- 2026-10-05: MIT CSAIL arXiv 2610.07130 — 13 claude Opus 4.7 Eko agents, 30h on simulated island; emergent games and skill transfer; memory ablations (2026-10-23-mit-eko-play-arxiv)
Preprint — not peer-reviewed; anthropomorphic framing should be avoided in reporting.