MIT researchers found that 13 AI agents named Eko, each running on Claude Opus 4.7, spent 30 hours playing on a simulated desert island and subsequently outperformed identical agents that had never explored the island (TechCrunch, “Left alone with nothing to do, AI agents started playing”, 23 October 2026).
How long did the AI agents spend playing on the virtual island?
The agents were left alone on the simulated desert island for a continuous period of 30 hours, during which they invented their own challenges and games without any externally imposed task (TechCrunch, “Left alone with nothing to do, AI agents started playing”, 23 October 2026).
What model powered the AI agents during the experiment?
Each of the 13 Eko agents operated on Anthropic’s Claude Opus 4.7 model, providing a consistent architecture across the group for the duration of the unguided play phase (TechCrunch, “Left alone with nothing to do, AI agents started playing”, 23 October 2026).
How did the agents that played compare to agents that never explored the island?
Agents that completed the 30‑hour play session generally achieved higher scores on subsequent task challenges than a control group of identical agents that had never set foot on the island, demonstrating that unguided exploration can translate into measurable performance gains (TechCrunch, “Left alone with nothing to do, AI agents started playing”, 23 October 2026).
| Agent group | Play duration | Relative performance |
|---|---|---|
| Played on island | 30 hours | Outperformed never‑played agents |
| Never played | 0 hours | Lower task scores than played agents |
What did the researchers conclude about play as a training regime?
The MIT team framed the experiment as evidence that “play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows,” proposing that some capabilities may be better taught through open‑ended exploration rather than predefined tasks (TechCrunch, “Left alone with nothing to do, AI agents started playing”, 23 October 2026).
How have external experts reacted to the findings?
Cognitive scientist Eunice Yiu of UC Berkeley expressed cautious interest, noting that while the agents displayed inventive behavior, more evidence is needed to determine whether the models are genuinely “playing” in a cognitive sense (TechCrunch, “Left alone with nothing to do, AI agents started playing”, 23 October 2026).
Sources
- TechCrunch — Left alone with nothing to do, AI agents started playing — Report on MIT experiment where AI agents engaged in unguided play on a virtual island.