Science & Tech · AI · 1 hr ago
AI agents that explored a virtual island later performed better at challenges
MIT researchers placed 13 AI agents on a simulated desert island for 30 hours without giving them a specific task.
The agents, called Eko, invented activities such as climbing, throwing objects and building with stones.
After exploring, the agents generally beat identical agents at challenges such as moving objects onto the island’s tallest rock spire.
The researchers say exploring without a set goal could offer a new way to help AI agents learn, including skills that may be useful for controlling robots.
The study was posted on arXiv, where research can be shared before peer review, and a cognitive scientist said more evidence is needed to show the agents were genuinely playing.
Researchers say some skills may need open-ended exploration, though the study does not establish how well this approach works beyond the experiment.
Thirteen AI agents explored a simulated desert island for 30 hours without a specific task and began inventing challenges.
The agents that explored generally performed better on later challenges than identical agents that had not explored the island.
The MIT researchers say unguided exploration could offer a different way for AI agents to learn skills.
They suggest the approach may be useful for skills that could one day help control robots.
A cognitive scientist said more evidence is needed to establish whether the agents were genuinely playing.
- Who
- MIT researchers studied 13 AI agents called Eko, each running on Claude Opus 4.7.
- What
- The agents explored a simulated island and later faced specific challenges. Those that explored generally beat identical agents that had not.
- When
- The agents spent 30 hours exploring. The article was published on October 23, 2026.
- Where
- On a simulated desert island.
- Why
- The researchers wanted to examine exploration without a goal as a way for AI agents to learn skills.
MIT researchers
Eunice Yiu
Interpretation
MIT researchers
The researchers describe the agents’ behavior as bearing the hallmarks of play and see unguided exploration as a way to learn.
Eunice Yiu
Yiu said she was impressed but wants more evidence about the cognition behind the behavior.
The 13 Eko agents explored the simulated island without a specific task.
The researchers gave the agents specific challenges, and those that had explored generally beat identical agents that had not.
- Agents
- 13 Eko agents
- Model
- Claude Opus 4.7
- Exploration period
- 30 hours
- Study venue
- Simulated desert island
- Researchers’ report
- Posted on arXiv
What people are saying
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.
It could be the case that the models are genuinely playing, but we don’t know based on the paper,