MIT's Eko Agent Taught Itself Island Skills Through 30 Hours of Unsupervised Play
Researchers left Claude Code on a simulated island for 30 hours with no task, and the agent spontaneously invented sports, built towers, and learned skills that transferred to later tests.
- MIT researchers left Claude Opus 4.7 on a simulated island for 30 hours with no task or reward.
- The agent spontaneously invented sports, built towers, drew mandalas, and ran physics experiments across 13 independent runs.
- Behaviors matched four classical signatures of play: self-created rules, repetition with variation, make-believe, and goals held lightly.
- Played agents outperformed fresh agents on four one-hour tasks including tower-building and Spire climbing.
- Erasing specific techniques from memory files causally removed downstream capabilities, proving in-context skill transfer.
- Code released: Clawblox engine, Eko harness, replays at machineplaying.org.
How Eko learned island skills through 30 hours of play
MIT researchers placed a coding agent in a small 3D world, gave it a body controlled through terminal commands, and removed any assigned objective or success reward. Across 30-hour runs, the agent, named Eko, created games, tested the island’s physics, drew patterns, and developed techniques that later improved its performance on measurable tasks. The study, titled Is This Machine Playing?, argues that these behaviors meet classical criteria for play and show how self-directed activity can support machine learning through persistent memory.
Why an empty task list matters
Most software-agent research begins with a defined task, an evaluation metric, and some form of feedback. Eko tests whether a capable agent can generate its own activities and acquire useful skills during otherwise unstructured runtime. For developers, the experiment isolates a practical learning channel: agent-written files can preserve procedures across sessions without retraining the underlying model.
Inside Eko’s island
Eko combines Claude Code, backed by Claude Opus 4.7, with a lightweight simulation harness. The island contains mesas, cubes, a rock, and a yellow ring governed by rigid-body physics, allowing objects to fall, collide, stack, and travel when thrown.
| Component | Implementation |
|---|---|
| Actions | MoveTo(position), Jump, Stop, Pickup(position), Place(position, rotation), and Throw(direction, speed) |
| Observations | A text dictionary containing each entity’s position, shape, and RGB color; the agent receives no images |
| Persona | SOUL.md supplies a curiosity-oriented persona |
| Episodic memory | EPISODIC_MEMORY.jsonl records events over time |
| Semantic memory | SEMANTIC_MEMORY.md stores durable, verified knowledge |
| Reset schedule | The Claude Code session and island reset every hour, while workspace files and programs persist |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
