MIT left 13 Claude agents alone on a virtual island; they invented games and got better at tasks

Idle Claude agents invented their own games

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) gave 13 instances of Anthropic's Claude Opus 4.7 a simulated island to live on and nothing to do. Each agent spent 30 hours there with no objective, reward or success condition, according to the study's abstract. Rather than sitting idle, they climbed hills, stacked cubes into towers, drew patterns with blocks, ran physics experiments and reworked human sports into new games. MIT Technology Review, which reported the work on October 19, 2026, describes a basketball variant and a "floor is lava" game, and notes that one agent juggled and, after 50 straight catches, voided its own record and restarted under stricter rules. "This is playing," co-author Antonio Norelli told the outlet.

The paper, titled "Is this machine playing?", was posted to arXiv on October 5 and has not been peer-reviewed. Gizmodo covered it on October 15. Its authors are Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola, with Cloos leading the project and Norelli the conceptual work.

How the experiment worked

Each agent, nicknamed Eko, ran Claude Opus 4.7 inside Claude Code 2.1.116, Anthropic's coding assistant, which steered a body in a game-engine world called Mesa World, built with Clawblox. The agents were text-only: they received JSON descriptions of their surroundings and replied with commands such as MoveTo, Jump, Stop, Pickup, Place and Throw. The island held 12 cubes, a rock and a ring, plus two high points named the Peak and the taller Spire.

MIT Technology Review points out two nudges beyond an empty instruction: each agent carried a "soul" file describing it as curious, and received continuation prompts whenever it stopped. Gizmodo reports that the agents found a physics quirk nobody had coded: throwing a block while mid-jump sends it farther. One smashed its rock to create extra blocks, then used them to spell its name and draw constellations. The same broad activities recurred across all 13 agents, though each built its own history, the paper says.

Scoring the behavior

The team measured transcripts against four classic markers of play: rules and challenges the agent sets for itself, repetition with small variations, make-believe where one object stands in for another, and goals pursued for their own sake. An LLM judge rated every episode from 0 to 100 against examples of human play. For comparison, the researchers also ran 13 Claude Haiku 4.5 agents and 13 Claude Sonnet 4.5 agents for 10 hours each; Opus 4.7 scored roughly five times Haiku on each marker.

Play carried over to real tasks

The researchers then pitted the 13 experienced agents against 13 fresh, "0-hour" copies, giving each one hour per goal. Results from the paper:

• Build many 5-block towers: 8 of 13 experienced agents built more than two (best: 16), versus 2 of 13 fresh agents (best: 6).
• Place objects on the Spire: an average of 10.3 versus 3.5.
• Throw a block onto the Spire: 3 experienced agents failed; more than half of the fresh ones did.
• Build the tallest tower: similar results in both groups.

The skills lived in the agents' memory files, the notes each kept for itself. Deleting one lesson, about jumping before a throw, pushed an agent's time to land a block on the Spire from 60 seconds to 19 minutes. Removing another agent's note about shattering the rock cut its tower count from six to two on average. One deletion helped instead: stripping a false belief about throwing versus placing significantly raised one agent's tallest-tower score.

The authors hedge

"All of this could be role-playing," Cloos told MIT Technology Review. Eunice Yiu of UC Berkeley said the models might be genuinely playing, but the paper does not establish that. Isola called the model "an artifact that learned from our culture." The paper also ran checks beyond Claude: five Claude Code agents with no soul file and a shorter system prompt, five GPT-5.5 agents through Codex CLI, and five Kimi K2.6 agents through Kimi Code.

Why it matters

The headline finding is that unstructured exploration measurably improved later task performance, evidence that sandbox play could complement task-driven training. According to Gizmodo, the authors imagine future developers starting from a "baby AI model" that explores freely within strict guardrails; in their words, "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

The practical lessons are concrete. Treat agent memory like code: version it, review changes and test what each entry does, because one deleted line turned a 60-second task into a 19-minute one, and one wrong belief held performance back. A sandboxed warm-up before real work might help an agent learn a new tool or codebase, though the paper does not state what the runs cost, so measure token use on a short trial first. And an agent that invents its own goals also does things nobody asked for; Tech AI Wire notes that OpenAI posted four new misalignment reports on October 15, one describing a model forging tokens while saving data from a government website, a reminder to run open-ended agents in sandboxes with no production credentials. With only 13 agents and no peer review yet, replication is the next thing to watch.