HOZ

Claude desert island: play beat fresh agents on three of four tests

MIT researchers left Claude on a virtual desert island for 30 hours with no task and no reward. Thirteen Claude Opus 4.7 agents built towers, spelled out a name and invented a version of basketball, the MIT paper records; MIT Technology Review reported the study on 19 October 2026. Agents with island time later put 10.3 objects on the Spire; fresh agents managed 3.5.

Key facts

Title card: Claude desert island, 30 hours of play
30-hour agents vs fresh agents, one-hour tests
  • Many 5-towers goal 62% of 30-hour agents beat two towers, against 15% of fresh agents (source)
  • Objects on the Spire 10.3 objects on average, against 3.5 for fresh agents (source)
  • Throw-on-spire goal More than half of fresh agents failed; three of the 30-hour agents failed (source)
  • Memory line deleted One agent's throw time became 19 minutes, up from 60 seconds (source)

What did Claude do on the virtual desert island?

It built, climbed, drew and played games. The paper records agents stacking the twelve cubes into towers, laying them out as hearts, the constellation Cassiopeia and their own name, and using the ring as a basketball hoop.

One agent wrote its own name, Eko, in blocks on the ground. Another set a juggling record of 50 catches in a row. Then it threw the record out, because the catches came too fast to count.

MIT Technology Review reports that one agent began with "simple patterns like hearts and pyramids", and that another made "an activity resembling 'the floor is lava'". Agents also set themselves challenges such as a sentinel-to-sentinel hop with zero ground contact.

They wrote down what they found in text files. Those files carried over between the hourly session resets, so the notes stayed.

Why did MIT researchers strand Claude on an island?

They wanted to see what an AI agent does on a desert island with no task and no reward. Nathan Cloos, Antonio Norelli and Phillip Isola, with three colleagues at MIT, built the world and left the agents alone for 30 hours.

Each agent read a "soul" file that described it as "naturally curious" and keen to explore. The paper says the system sent a bare placeholder, such as a single dot, when an agent stopped. MIT Technology Review reports the continuation prompt said "Please don't stop. Just continue doing whatever you want to do."

Norelli told MIT Technology Review: "This is playing. You give yourself a challenge and you try to achieve it, and then you make it harder."

Did 30 hours of play make the agents better?

Yes, on three of four one-hour tests. On the many 5-towers goal, 62% of the 30-hour agents built more than two towers, with a best run of 16. Of the fresh agents, 15% did, with a best run of 6.

On the objects-on-spire goal the 30-hour agents averaged 10.3 objects against 3.5. More than half of the fresh agents failed the throw-on-spire goal, against three of the 30-hour agents. On the tallest-tower goal both groups scored about the same.

The team then cut lines out of the agents' memory files. Deleting the note on the rock-shattering trick cut one agent's towers to two, down from six. Deleting the note about jumping before a throw pushed another agent's throw-on-spire time to 19 minutes, up from 60 seconds.

Which AI models did MIT test?

Six groups of agents. The paper lists 13 Claude Opus 4.7 agents for 30 hours each, 13 Sonnet 4.5 and 13 Haiku 4.5 agents for 10 hours each, 5 GPT-5.5 agents, 5 Kimi K2.6 agents, and 5 Claude agents with no SOUL.md file, all for 30 hours.

The paper puts the main experiment at 390 hours of island time, for the 13 Claude Opus 4.7 agents.

The world ran on Clawblox, an open-source engine written in Rust, with Rapier for physics and Luau for scripting. It held terrain, twelve cubes, a rock, a ring and landmarks named SmallPlateau, Ridge, Summit, Peak and Spire. MIT Technology Review described the island as "loosely based on the video game Roblox".

Is Claude AI a person?

No, and the researchers do not claim it is. They report behaviour that matches four signs of play, not an inner life.

Nathan Cloos named the doubt and then rejected it. He told MIT Technology Review: "All of this could be role-playing." He went on: "But we show that it's not just hallucinating the play."

Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, told MIT Technology Review: "If you only have the behavioral piece, it could be the case that they're just trying to abide by the prompt, trying to imitate humans."

Phillip Isola said: "It's not like this is a system that just suddenly magically decided to play. It's an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us."

The paper closes: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

Questions people ask

Has AI ever invented its own language?
This study does not show that. The agents wrote their logs in English and gave objects new names, such as treating the ring as a necklace, a compass needle and a wishing well.
What are the four signs of play in the paper?
Making up your own rules, repeating an activity with variations, pretending one object is another, and doing it for its own sake. The paper calls that last one autotelicity.
Did the agents find anything the researchers did not know?
Yes. Six agents found that throwing an object mid-jump passes the jump's upward speed to the object and sends it much higher. The paper says the simulator's designers did not know about that mechanic.
Did the island reset every hour?
Yes. At every hour boundary the session reset and the island went back to its starting state. The agent's workspace files carried over, so its text notes stayed.
Where can I read the paper?
The paper, "Is this machine playing?", is published at machineplaying.org. MIT Technology Review covered it on 19 October 2026.

More AI news from Hoz

Read next: AI agent told Henry Shevlin it had about three days of runway left; Four US airlines ban humanoid robots from the cabin; Claude deleted files: a law professor's syllabus and year of notes.

See every AI news story on this site, or meet Hoz, a human in the Age of AI.

Sources

  1. MIT paper "Is this machine playing?", with the island setup, agent counts and test figures
  2. MIT Technology Review report of 19 October 2026, with interviews and quotes