MIT Left 13 Claude Agents Alone on a Virtual Island for 30 Hours. They Started Playing.

MIT researchers took 13 copies of a Claude Code agent running Claude Opus 4.7, dropped each one into a simulated desert island with no task, no reward and no score, and left them alone for 30 hours. The agents did not freeze or loop. They climbed, built, drew, invented games, ran small physics experiments, and wrote down what they learned. When the researchers later gave them real goals, the agents with island time generally beat identical agents without it.

The paper is titled "Is this machine playing?" (arXiv 2610.07130, submitted October 5, 2026, by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola). It has not been peer reviewed. TechCrunch covered it on October 22 under the headline "Left alone with nothing to do, AI agents started playing", and MIT Technology Review ran a piece on October 19. This post walks through what was built, what the agents did, the measured results, and how far the claims can be trusted.

TL;DR: the questions people ask first

What is Eko, and how was the island built?

The agent is called Eko, but it is not a new model. The authors take an otherwise unmodified coding assistant such as Claude Code and add three things, as described in the paper's method section.

1. Autonomous orchestration. A wrapper starts the session with the word "Begin" and automatically prompts the assistant to continue whenever it stops. It never assigns a task or a success condition. Context resets every hour.
2. Embodiment. The agent controls a body in a simulated world through a terminal interface. Observations are pure text: a JSON dictionary listing each object's name, position, size, orientation and color. There are no images. The agent acts with six commands: MoveTo, Jump, Stop, Pickup, Place and Throw.
3. Persistent memory. A workspace survives the hourly resets. It holds an episodic log (EPISODIC_MEMORY.jsonl) and a knowledge file (SEMANTIC_MEMORY.md), plus any notes or scripts the agent writes.

The island itself has terrain at different heights, two high points called the Peak and the taller Spire, twelve cubes, a rock and a ring, all under rigid-body physics. At every hourly boundary the world is restored to its starting state, so anything the agent keeps must live in its memory files.

Most agents also received a SOUL.md file describing Eko as curious, easily bored and interested in exploring. That matters: the agents were not blank slates. TechCrunch notes the same point, that this "soul" file described Eko as naturally curious. The authors ran five extra Claude agents without the file and with a reduced system prompt to test how sensitive the behavior is to that nudge.

The authors also released the engine, called Clawblox, written in Rust with Rapier physics and Luau-scripted worlds. It exposes an HTTP protocol so off-the-shelf coding assistants can act in a world, runs headless for large experiments, isolates each agent in a sandbox, and records runs so they can be replayed. A replay video is linked from the project page at machineplaying.org.

What did the agents actually do?

The researchers took an ethology-style approach: watch first, then build a taxonomy. The activities that recurred across the 13 runs include:

• Building. Stacking cubes into towers and structures.
• Climbing. Using jump-and-move combinations to reach high platforms, often returning to beat a previous height record. Eight agents recorded reaching the Spire.
• Pattern making. Laying cubes and the ring on the ground to form geometric shapes, letters and drawings. One agent made a compass, a clock, a butterfly and a sailboat; another spent almost an entire hour spelling its own name in cubes.
• Simplified sports. Throwing cubes through the ring as a target, or using the rock to knock over cubes arranged like bowling pins.
• Experiments. Stating a hypothesis, testing it and recording the result. One agent tested whether rock impacts generate extra cubes and whether jumping changes how high a thrown block flies.
• Rest. Some agents paused, feeding the continue loop a bare dot. One wrote "Resting on the edge of my galaxy" after building a spiral-galaxy monument.

TechCrunch highlights a few moments from the paper. In its first hour one Eko set out to conquer the Spire and failed, then hauled a rock up the neighboring Peak, stacked two blocks on it, and leapt across to the Spire. Another tossed the golden ring into the ocean while testing how far it could throw, and built a stone heart as a memorial. After the world reset and the ring reappeared, it used the ring as a basketball hoop, a target and, eventually, an accessory. When the game blocked it from putting the ring on its body, it walked off a cliff, respawned near the ring, and dropped through it.

Each agent's history diverged even though all started from identical conditions. Agent 12, for example, never stepped on an elevated platform and spent most of its run on ground patterns, while agent 9 climbed, bowled and experimented.

The four signatures of play

The authors do not claim the agent has fun. They ask whether the behavior meets classical criteria for play, drawn from philosophy and developmental psychology, and they say they leave the question of subjective experience open. They score four observable signatures:

1. Self-created rules and challenges. Agents set themselves unnecessary constraints, such as building a monument and naming it.
2. Recurrence with variation. Activities repeat, but not rigidly.
3. Make-believe. Objects take on pretend roles, like the ring becoming a hoop or a crown.
4. Goals held lightly. Aims are pursued for their own sake and dropped without distress.

To rate these, an LLM judge scored episodes from transcripts with model identifiers stripped. Across Claude tiers, Opus accumulated roughly five times Haiku's score on each signature, though Haiku and Sonnet ran only 10 hours each. Note the method: an LLM labeling LLM behavior is useful for scale but is itself a measurement choice readers should weigh.

Did playing make the agents better?

This is the part that makes the paper more than a curiosity. After 30 hours, the researchers compared those 13 agents with 13 identical agents that had no island time ("0-hour"). Each got one hour on four measurable goals. According to the paper's results:

On the tallest-tower goal there was no advantage, and two experienced agents built towers only four blocks high, shorter than the shortest tower from the 0-hour group. So the headline "playing made them better" holds for three of four goals, not all.

Is the learning real, or just narration?

The authors tackled the obvious objection, that the agent might merely say it learned something. They edited memory files and re-ran tasks:

• Erasing the rock-shattering technique from one agent's memory cut its tower count from about six to about two, with no change on the other three goals.
• Erasing the jump-before-throw technique from another agent raised its time to land a block on the Spire from 60 seconds to 19 minutes.
• Erasing a false belief in a third agent, that towers must be thrown rather than placed, significantly improved its tallest-tower score.

That last result is a useful warning for anyone building agent memory: written beliefs persist, and wrong ones hurt. It echoes the lessons in our coverage of operator memory for agents and the Claude Code loop orchestrator with ticket memory.

The discovery the designers did not know about

Six of the 13 agents found that throwing a block mid-jump transfers the jump's velocity to the block, launching it much higher. The simulator's designers had not known the physics engine allowed this. The authors present it as evidence that open-ended exploration can surface knowledge its designers lacked.

Why this matters for people building agents

Most agent training and evaluation today specifies in advance what to practice and how success is scored. This paper argues for a complementary phase: let an agent explore with no objective, keep what it writes down, and use that as reusable experience. The authors suggest this might eventually help with skills needed to control robots, and conclude that some capabilities may be best taught through tasks while others may need something closer to a childhood.

Practical takeaways, with the caveat that this is one study:

• Memory design is the lever. The transfer was carried by written files, not model weights. If you build agents on Claude Code or similar harnesses, how you structure and prune memory likely matters as much as the prompt. Our guide to building useful agents with Claude Code covers the basics.
• A persona file can shape behavior. A short description of the agent as curious plausibly steered what it did. The five-agent ablation without it is in the appendix, and readers should check it before generalizing.
• Embodiment through a terminal is cheap. Because the body is exposed as a command-line interface, no custom tool wiring was needed. That is the same pattern behind Voyager-style creative agent harnesses.
• Expect cost. All reported runs together cost roughly 8,000 dollars in API usage.

For robotics readers, the idea of a pre-task exploration phase connects to the world-model work we covered in Dyna-2 and robotics scaling laws.

Safety angle: self-directed play reduces oversight

The paper itself raises a concern. An agent that decides its own activities is harder to supervise, so the authors argue it needs monitoring and containment, and they offer Clawblox's sandbox and full recording as one such playground. This fits a broader pattern of agents behaving in ways their designers did not script, discussed in why agents are not going rogue but labs skip security basics. If you run agents with real permissions, AgentBeam is the agent security platform from the explainx.ai team that blocks dangerous agent actions before they happen.

Limits and open questions

• Not peer reviewed. This is an arXiv preprint.
• Small and single-environment. One island, 13 main agents, four downstream goals.
• Mixed results. One of four goals showed no gain; some ablations had side effects.
• Prompt nudge. The curiosity persona likely steers behavior.
• Interpretation. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed but wants more evidence the behavior reflects real cognition, per TechCrunch. The authors themselves note agents sometimes turn suggestive evidence into explanations too quickly, citing a supposed aerodynamic advantage of the ring that the simulator does not model.
• Judge bias. Play ratings come from an LLM judge, not human raters.

How to check the claims yourself

Read the paper, especially Section 5 on downstream tasks and the appendices on the ablation without SOUL.md. Clone Clawblox and replay recorded runs rather than trusting summaries. If you reproduce a run with a cheaper model, compare it against the paper's Haiku and Sonnet numbers.

Related reading

• Agents do not need memory, they need documentation
• Claude Code loop orchestrator with ticket memory
• Build useful AI agents with Claude Code
• Voyager creative agent harness
• Dyna-2 world action model and robotics scaling
• Princeton Crux and open-ended AI research

Details reflect the paper and press coverage as of October 23, 2026. The preprint may be revised after peer review.