Claude Was Left Alone on an Island for 30 Hours. It Started Playing

Researchers at MIT CSAIL, the Computer Science and Artificial Intelligence Laboratory at MIT, left a copy of Claude Opus 4.7, an Anthropic AI model nicknamed Eko, alone on a virtual island for 30 hours. They repeated this several times. The agent was only given a personality description of a curious explorer, with no other instructions. The agent invented basketball and juggling on its own, working its way up to 50 consecutive shots, then made its own rules harder. When the agents were later tested on a separate task, carrying as many objects as possible to the island's highest point, the copies that had spent 30 hours playing outperformed fresh copies with no such experience.

This points to a third way of training agents, alongside fine-tuning (further training a model on new examples after its main training) and reinforcement learning (where a model learns by trial and error while chasing a reward). The agent gets better simply because it invents its own activities and keeps making them harder on its own. For practice with Claude Code, Anthropic's coding tool, this is not a usable tool yet. It is just a signal. The model has a reserve of self-directed behavior that does not show up during normal work on preset tasks. No one should expect this to land in a product tomorrow. But anyone planning long autonomous agent runs now has a reason to watch what the agent does during idle time, not just when it is given an explicit task.