Claude Opus 4.7 agents invent games on a virtual island
Researchers at MIT left 13 copies of Anthropic's Claude Opus 4.7 alone on a virtual island for 30 hours each, with no task to complete. The agents invented games, ran physics experiments and drew pictures with blocks. Afterward, they beat fresh copies of the same model on new tasks. That suggests f
Posted on Oct 20 • Originally published at techaiwire.com
Researchers at MIT left 13 copies of Anthropic's Claude Opus 4.7 alone on a virtual island for 30 hours each, with no task to complete. The agents invented games, ran physics experiments and drew pictures with blocks. Afterward, they beat fresh copies of the same model on new tasks. That suggests free exploration time can make an AI agent better at real work.
MIT Technology Review reported the study on October 19, 2026. Gizmodo covered it on October 15. The paper, titled "Is this machine playing?", was posted to arXiv on October 5 and has not been peer-reviewed yet.
The authors are Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola. Cloos led the project and Norelli led the conceptual work, the paper says. Gizmodo and MIT Technology Review place the team at MIT CSAIL, the university's computer science and AI lab.
Each agent ran Claude Opus 4.7 inside Claude Code version 2.1.116. Claude Code is Anthropic's coding assistant, which can run commands on its own. Here it steered a body on a simulated island called Mesa World, built with a game engine called Clawblox. Each agent was nicknamed Eko.
The agents never saw pictures. They received text descriptions of the world in JSON, a structured data format, and replied with commands.
The instruction named no task, reward or activity, according to the abstract. MIT Technology Review points out two nudges, though. Each agent had a "soul" file that described it as curious. It also received "continuation prompts" whenever it stopped.
With nothing to do, the agents climbed hills, stacked blocks into towers, drew mandalas and reinterpreted sports, the abstract says. MIT Technology Review describes a basketball variant and a "floor is lava" game. One agent smashed a rock to make extra blocks. It then used the blocks to draw its own name and constellations.
Another agent juggled. After 50 catches in a row, it voided its own record and restarted under stricter rules. "This is playing," Norelli told MIT Technology Review.
Gizmodo reports that the agents found a trick the designers had not coded. Throwing a block while mid-jump sent it farther. The activities recurred across all 13 agents, but each agent built its own history, the paper says.
The team scored the behavior against four classic signs of play:
An LLM judge, meaning another AI model grading the transcripts, rated each episode from 0 to 100 against examples of human play. The team also ran 13 Claude Haiku 4.5 agents and 13 Claude Sonnet 4.5 agents for 10 hours each. Opus 4.7 built up about five times Haiku 4.5's score on each sign.
On most tests, yes. The researchers compared 13 agents with 30 hours of island experience against 13 fresh "0-hour" agents. Each agent got one hour per goal.
The skills lived in the agents' memory files, the notes each agent kept for itself. The team deleted single lessons from those files and ran the tests again. One agent lost its note about jumping before a throw. Its time to land a block on the Spire rose from 60 seconds to 19 minutes. Another lost its note about shattering the rock. Its tower count fell "from six to two on average."
One deletion helped instead. Removing a false belief about throwing versus placing "significantly increased" one agent's tallest-tower score, the paper reports.
The authors stay cautious. "All of this could be role-playing," Cloos told MIT Technology Review. Eunice Yiu of UC Berkeley said: "It could be the case that the models are genuinely playing, but we don't know based on the paper." Isola called the model "an artifact that learned from our culture."
The paper probes that worry in part. It ran five more Claude Code agents with no soul file and a shorter system prompt. It also ran five agents on GPT-5.5 through Codex CLI and five on Kimi K2.6 through Kimi Code.
Gizmodo reports the authors' wider idea. Future developers could start with a "baby AI model" that explores freely, inside strict guardrails. In the researchers' words: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."
This article was first published on Tech AI Wire.
Deutsch · 日本語 · Français · Español · Português