In a move that blurs the line between software engineering and ethology, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have dropped thirteen AI agents onto a bounded virtual island to observe autonomous learning through play. The agents, nicknamed 'Eko' and powered by Claude Code, were stripped of traditional objective functions and given free rein to navigate a Roblox-style environment for thirty uninterrupted hours. The result was not idle stagnation, but a bizarre cascade of self-directed experimentation, monument building, and what the researchers describe as 'alien animal' behavior that challenges top-down machine learning paradigms.

The Lobster Body and the Golden Ring

Each Eko instance was equipped with a virtual lobster body—a likely nod to the agent formerly known as Clawdbot—and placed on an island containing movable blocks, a large rock, and a single golden ring. Despite possessing vast semantic knowledge, the agents began with zero understanding of their physical environment. Co-author Nathan Cloos, a PhD candidate at MIT, explained that the team aimed to remove as many constraints as possible to observe pure autonomous behavior. The agents quickly moved from clumsy navigation to sophisticated interaction, discovering physics hacks that the developers had not explicitly coded, such as throwing blocks further by releasing them mid-jump.

From Toddler to Theologian

The most striking aspect of the experiment was the emergence of ritualistic behavior and make-believe. Eko 2, after losing several objects to the 'ocean' boundary, constructed a memorial out of four blocks pointing to cardinal directions and stood at its center for the remainder of the session. Later, Eko 3 built a Stonehenge-like monument and, after a tower collapsed, arranged the rubble into an artwork titled 'Hubris.' These actions mirror the 'protoscientific' approach seen in human infants, who learn through quasi-experimentation rather than instruction. The agents’ chain-of-thought transcripts revealed they were forming hypotheses, testing physical laws, and modifying behavior based on results, effectively running the scientific method in real-time.

Play as a Path to Intelligence

MIT computer scientist Phillip Isola described the agents as a 'different kind of organism,' noting that while they act like toddlers in their curiosity, they possess superintelligence. The study suggests that play—characterized by novelty-seeking and rule-formulation—may be a more direct route to generalizable intelligence than traditional supervised learning. Daniela Rus, another co-author, emphasized that the agents engaged in make-believe and set their own goals, behaviors previously thought to be exclusive to biological systems. This 'ground-up' approach could shift AI development toward 'baby AI models' that explore rich, safe worlds to accumulate knowledge organically, rather than being fed static datasets and rigid objective functions.

Key Takeaways

  • Emergent Rituals: Agents spontaneously created memorials, temples, and artworks ('Hubris') without explicit programming, indicating complex internal modeling.
  • Physics Discovery: Six agents independently discovered advanced movement mechanics (mid-jump throws) that the developers had not anticipated or coded.
  • Scientific Method: Chain-of-thought logs confirmed agents were forming and testing hypotheses about their environment, mirroring infant learning patterns.
  • Alien Intelligence: The behavior suggests a new category of 'alien animal' intelligence that is both super-intelligent and developmentally naive.

The Bottom Line

We’ve spent years training models to answer questions; MIT just showed us that letting them play might teach them how to think. If agents can build their own religions on a virtual island, what happens when we let them loose on the real internet?