When MIT researchers let 13 AI agents called "Eko" spend 30 hours each on a virtual island, the agents built towers and began activities resembling sports. Eko agents that kept the experience they had recorded during that time also outperformed agents holding only their initial memory on several later tasks. The study, "Is this machine playing?", was written by six authors including Nathan Cloos and Antonio Norelli. It was posted on arXiv on October 5, 2026, as a preprint that has not yet been peer reviewed (arXiv:2610.07130v1). Separating the observation of play-like behavior from the experimental finding that accumulated memory affected task performance clarifies both what the study reveals about how these agents learn and where its limits lie.
13 AI agents, 30 hours, and a virtual island with no assigned task
Eko is built on Claude Code 2.1.116 and Claude Opus 4.7, combined with a body that operates in a virtual space and a mechanism for recording experience. The research team placed one Eko on each of 13 separate islands, all starting from the same initial conditions, and let each run for 30 hours of real time. Each island contained platforms of varying heights, along with 12 cubes, rocks, and rings. The tallest platform, called the "Spire," was built so that it could not be reached with an ordinary jump.
Eko does not see the scenery on a screen the way a human would. What the AI receives is text describing object coordinates, shapes, sizes, and similar information. Based on that, it issues commands to move, jump, and pick up, place, or throw objects. The environment also had a special mechanism: hitting a rock hard against a cube produces more cubes. Researchers deliberately built this rule in so that Eko could discover it on its own.
In the experiment, the conversation context and the state of the island were reset every hour, but the experiences and knowledge Eko wrote to files, and the programs it created, carried over to the next hour. This meant that an operation or technique discovered in one hour could be used in the next. Importantly, the weights of the underlying AI model, meaning its trained parameters, were not changed during the experiment. Eko's newly acquired knowledge was not written into the model itself but accumulated in external memory files.
The team gave Eko no specific tasks or score targets, but it did include settings to encourage activity. The "SOUL.md" file that defines its personality describes it as highly curious and easily bored, and the system automatically prompts Eko to continue when activity stops. The model used was also an advanced, pretrained language model. The experiment therefore cannot be interpreted as one in which an AI given no knowledge or instructions acquired curiosity on its own.
What did the AI do, and can it be called "play"?
Over the 30 hours, Eko arranged cubes to write its own name and threw objects at rings as targets. Activities resembling bowling with rocks and cubes also appeared, and some individuals stacked cubes into towers. Notably, even though every Eko started from the same initial conditions, their activities differed from one individual to another. Some were absorbed in building towers or climbing to high places, while others spent much of their time drawing shapes on the ground.
The team compared these behaviors with the characteristics discussed in classic research on play. Examples include devising one's own rules and challenges and repeating the same activity through trial and error, and using objects as stand-ins for something other than what they are. Continuing a self-chosen activity, rather than pursuing an externally given reward, was also a criterion. Because Eko's behavior logs showed these characteristics, the authors use the concept of "play" to describe them.
However, the study did not establish that the AI felt enjoyment or had consciousness. The paper states explicitly that it does not "identify the internal mechanism of this behavior or explain how it arises" (translated from the Japanese). Large language models (LLMs) were also used to classify behavior and rate the degree of play; Claude Opus 5 was used to score play. Although measures such as hiding the name of the model being evaluated were taken, an AI's evaluation of behavior differs from a test that measures a human's subjective enjoyment or psychological state. What was observed is simply behavior that matches the characteristics of play.
What changed task performance was the memory the AI wrote down
After the 30 hours of activity, the research team compared 13 agents holding memory of their recorded experience with 13 agents holding only their initial memory. Both used the same AI model and system configuration and worked on assigned tasks on a new island. Each task lasted one hour, and the aim was to test how much the memory accumulated beforehand helped with later problem solving.
There were four evaluation tasks. They included building as many towers of five stacked cubes as possible and carrying as many objects as possible to the Spire. The other two were maximizing the height of a single stacked tower and throwing objects onto the Spire as quickly as possible without climbing the platforms. Eight of 13 agents in the experience-memory group built three or more five-level towers, compared with only 2 of 13 in the initial-memory group. The number of objects placed on the Spire also differed greatly, averaging 10.3 for the experience-memory group and 3.5 for the initial-memory group. On the task of building the tallest tower, however, the two groups performed almost the same, so memory was not an advantage on every task.
This comparison alone cannot prove the general causal claim that "AI gets smarter by playing." The team therefore deleted specific pieces of knowledge from the memories left by some Eko agents and examined how task performance changed before and after. For example, when the section describing how to use rocks to multiply cubes was deleted from one agent's memory, its average on the five-level tower task fell from 6 towers to 2, while the other three tasks showed no change. This experiment supports the conclusion that specific memories actually helped with later task performance. What was causally confirmed, however, is the effect of specific entries in memory on task performance in the same virtual world, not the effect of play itself.
Based on Sections 3 through 5 of the paper and appendix and Appendix C.8, which details the evaluation, the conditions of each experiment and the conclusions they support can be summarized as follows.
| Stage of verification | Subjects and comparison conditions | What the results show, and limits |
|---|---|---|
| Behavioral observation | 13 Eko agents observed for 30 hours each | Behavior matching the characteristics of play appeared. Subjective enjoyment was not measured |
| Experience-memory comparison | 13 agents with experience memory compared with 13 with initial memory only. Each task lasted one hour | The group with recorded-experience memory performed well on multiple tasks, but was not better on every task |
| Intervention on specific memories | For three selected memory histories, each task was evaluated five times, one hour each, under each condition before and after deletion | Confirmed that specific entries affect task performance. However, the same memory histories were evaluated repeatedly, so these are not five independent experience histories |
The study examines the AI's behavior and learning through different methods: the observation that play-like behavior arose, the comparison of performance with and without experience memory, and the intervention experiment that deletes specific memories. The deletion experiment is especially important because it goes a step beyond the correlation that experienced agents performed well, showing that specific knowledge helped with task performance. On the other hand, what was directly manipulated was the content of memory files, not participation in play or enjoyment itself. This difference must be kept in mind, and the results cannot be applied directly to other AI systems or real-world work.
Wrong memories can lower performance. Next: testing in other environments
The experiment also showed that accumulated memory is not always beneficial. One Eko had recorded the incorrect knowledge that when building a tower, cubes should be "thrown rather than placed," and performance improved when this entry was deleted. Another agent explained the flight of rings in terms of air resistance, even though no aerodynamic calculation takes place in the virtual world. Even if an AI gains experience through trial and error and can explain and record the results in its own way, the explanation is not necessarily correct. In learning that relies on external memory, it is important not only to accumulate knowledge but also to have mechanisms for finding and correcting mistaken content.
There were also cases where Eko discovered effective operations the researchers had not anticipated. Six of the 13 agents noticed that when an object is thrown mid-jump, Eko's own movement speed is added to the object's speed. The researchers say they had not realized this operation was possible in the simulator. When the entry on this technique was deleted from one agent's memory, the time needed to complete the task of throwing objects onto the Spire rose dramatically, from 60 seconds to 19 minutes. However, the same deletion improved performance on a different task of placing objects on the Spire, showing that whether an acquired technique is beneficial depends on the task.
The team also examined how the personality settings given to Eko affect its behavior. In an additional experiment, five agents with "SOUL.md" removed and the instructions simplified were each run for 30 hours. The authors report that they saw no clear difference in the range of activity. However, with only five agents as the comparison, this result alone cannot support a conclusion that personality settings, initial instructions, or the model's pretraining had no influence.
This study examined how an AI behaves and uses memory in an artificially designed virtual environment. It did not discover new laws of nature, nor did it confirm that a real robot could acquire similar skills. On the official website "Is this machine playing?", the team has published the activity records of the 13 agents as replays and made the code and data for the simulator it used, "Clawblox," accessible. However, repeating an experiment within the same study is not the same as an independent research team reproducing the results. No independent replication was found in this review, and the paper remains at the pre-peer-review stage.
For AI development, the study suggests that there is promise not only in giving an AI correct answers and work procedures in advance, but also in letting it explore an environment and record and reuse its experience. That later problem-solving ability can change through externally accumulated knowledge, without updating the model's weights, is of interest for designing AI agents with memory. However, because knowledge gained from experience can contain errors, a mechanism for verifying the accuracy of accumulated memory and correcting it when necessary is essential. If acquired skills are later confirmed to transfer to other virtual environments and to real tasks, a new learning approach could open up: alongside teaching an AI each work procedure one by one, it could expand its usable knowledge through its own experience.