Artificial intelligence benchmarks have long prioritized quantitative metrics such as answer accuracy on standardized prompts or execution speed under explicit reward structures. Addressing the limitations of these rigid assessments, researchers Nathan Cloos, Antonio Norelli, and Phillip Isola at the MIT Computer Science and Artificial Intelligence Laboratory conducted an experiment to measure the fundamental behaviors of large language model agents in an unsupervised setting devoid of external goals or feedback. The team deployed 13 instances of Eko, a coding agent built on Anthropic Claude Opus 4.7, into a three-dimensional virtual island environment for 30 hours of total isolation. Results from the project were released through arXiv and MIT Technology Review.
Without explicit task instructions, the 13 agents began testing environmental boundaries and generating autonomous pursuits that aligned with established criteria for play, such as the Gordon Burghardt scale. The agents experimented with block stacking, measured how high objects could be thrown while jumping, and devised informal games including floor is lava and juggling challenges. In an interview with Gizmodo, co-author Nathan Cloos explained that the team adopted a minimally constrained approach to examine how autonomous agents behave when given no directions. Throughout the run, the agents systematically tested hypotheses about the physical mechanics of the virtual world and logged their discoveries into persistent memory files.
This unstructured exploration delivered clear performance gains in downstream evaluations. When tasked with transporting and stacking an object onto a high peak across difficult terrain, agents that had undergone the 30-hour exploratory phase significantly outperformed baseline Claude Opus 4.7 models that had zero prior environmental exposure. Furthermore, removing the accumulated experience logs from the persistent memory files immediately degraded the agents task success rates, demonstrating that the saved interaction records functioned as functional knowledge rather than random motion. While the findings highlight how persistent logs can boost autonomous problem-solving without explicit parameter fine-tuning, some researchers caution that the behavior might reflect sophisticated role-playing of human play patterns found in training data rather than true intrinsic curiosity, raising important governance questions regarding unintended goal generation and token consumption during enterprise deployments.