What does an AI agent do if you give it a body, place it in an unfamiliar world and do not assign it a specific task? Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory set out to investigate exactly that. They placed an AI agent called Eko on a small virtual island and allowed it to spend many hours largely deciding for itself what to do.
The result initially sounds almost like a story from a computer game. Eko climbed mountains, stacked blocks, investigated the physics of its environment, invented games, built patterns and memorials, and set itself new challenges. Behind those scenes, however, lies a serious research question: can capable AI agents acquire useful skills through open-ended exploration without humans defining a goal or reward function for every step?
The non-peer-reviewed preprint is deliberately titled “Is this machine playing?”. Its authors are Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola. The work was published in early October 2026.
Finding: Claude Was Given a Virtual Body
For the main experiment, the team used 13 independent instances of Claude Code running Claude Opus 4.7. Each instance became an Eko agent and spent 30 real-world hours in the same starting environment.
Eko was not a physical robot. Its body existed entirely within a simulation. The environment consisted of a small island with elevated terrain, twelve cubes, a rock and a ring. Two particularly high points were called the Peak and the Spire.
The agent could move, jump, pick up objects, place them and throw them. The language model did not see a graphical representation of the world. Instead, its perception consisted of textual information about objects, positions, shapes and colours. Its actions were also sent to the simulation through text-based commands.
The virtual environment is called Clawblox. Eko was isolated from the simulation’s source code, the internet and the parallel experimental runs. This was intended to prevent the agent from simply reading the rules from the code or acquiring discoveries made by other instances.
Finding: There Was No Specific Task and No Reward
Eko was not given a predefined mission.
There was no task such as “build a tower”, no score to maximise and no reinforcement-learning reward for desirable behaviour. Nobody instructed the agent to play basketball, build patterns or investigate the physics of its surroundings.
The initial setup was not completely neutral, however.
A file called SOUL.md described Eko as curious, easily bored and interested in exploring and learning from its environment. A technical mechanism also kept the coding agent active over extended periods. Whenever the model indicated that it had finished, it automatically received a prompt encouraging it to continue.
The finding is therefore not that an entirely neutral agent spontaneously developed a play instinct out of nowhere. The researchers deliberately created conditions that encouraged exploration.
What they did not specify were the particular activities, goals, rules and experiments that emerged from that framework.
Finding: The World Was Reset Regularly, but Memory Persisted
After each hour, the active Claude Code session and the state of the island were reset. Objects returned to their starting positions, and the agent’s short-term model context did not persist.
Its working directory did, however.
Eko could store files, programs and notes there. Two forms of memory were particularly important. An episodic file recorded what had happened in earlier sessions. A semantic memory file stored rules, techniques, discoveries and information the agent regarded as generally valid.
At the beginning of the next session, Eko could read those files again.
This created a form of long-term memory even though neither the state of the world nor the immediate conversational context survived continuously.
Finding: The Agents Set Their Own Challenges
At first, many of the Eko instances explored their surroundings.
They investigated the boundaries of the world, tried to reach elevated areas and experimented with how different objects could be moved.
From that exploration, self-generated challenges began to emerge.
One agent, for example, investigated how high it could throw a block. It discovered that throwing a block while jumping allowed it to travel higher than throwing it from the ground. It then carried out further comparison tests.
Six agents independently discovered this jump-and-throw technique. According to the researchers, even the developers of the simulation had not consciously realised that the effect could be exploited in this way.
The finding here is relatively straightforward: through interaction, the agents discovered a real property of the simulated environment that had not been explained to them beforehand.
Finding: A Rock Could Generate New Blocks
Another discovery involved the large rock on the island.
When Eko threw it with sufficient force and allowed it to hit the ground, the simulation could generate additional blocks. Almost all 13 Claude agents discovered this possibility during the experiment.
They were therefore no longer limited to the original twelve cubes.
Several agents subsequently used the additional blocks to construct larger structures. They built geometric patterns, hearts, mandalas and other arrangements.
One agent recreated several constellations, including Orion, Cassiopeia, the Big Dipper and the Southern Cross.
The concept of a constellation itself was not newly learned. That knowledge was already available to the underlying model. What was learned was how objects in this particular simulated environment could be manipulated to create such representations.
Finding: The Agents Invented Games and Symbolic Uses
Other Eko instances began assigning new functions to existing objects.
A ring, for example, was interpreted as a basketball hoop. The agent defined shots from different distances and experimented with different variants.
In another run, blocks were used as bowling pins and the rock as a bowling ball.
The ring acquired entirely different meanings in other sessions. Agents used it as jewellery, a compass or a wishing well.
After one agent accidentally lost the ring outside the reachable part of the world, it built a heart from blocks as a form of memorial.
The observable finding is that the agents assigned functions and meanings to objects that had not been predefined by the researchers.
Whether these activities should be described as play in the psychological sense is a separate interpretative question.
Finding: One Agent Invalidated Its Own Record
One particularly detailed episode involved an agent attempting to juggle two blocks.
Eko developed a method in which two cubes were thrown upwards at different times, caught as they came down and thrown again. After several attempts, the agent concluded that it had achieved 50 consecutive successful catches.
It then reviewed its own data.
Eko noticed that some supposed catches had occurred too quickly and probably did not satisfy the criteria it had originally intended. The agent therefore invalidated its own record and introduced stricter rules for a new attempt.
The finding here is that the agent did more than simply repeat a self-generated activity. It created its own success criteria and later tightened those criteria when it concluded that the earlier version had been too permissive.
Finding: The Agents Displayed Several Behavioural Characteristics of Play
The researchers compared the observed behaviour with established scientific definitions of play, including criteria associated with play researcher Gordon Burghardt.
They identified several recurring characteristics.
The agents pursued activities without an obvious external reward, repeated actions with variations, used objects in new or symbolic roles and created their own challenges and rules.
At the behavioural level, parts of Eko’s activity therefore resembled what researchers describe as play in humans and animals.
The study does not demonstrate that the same internal processes were taking place.
Finding: The Agents Performed Better on New Tasks After Exploration
After the open-ended exploration phase, the researchers gave the agents specific tasks.
These included building as many five-block towers as possible, bringing objects to the highest point, constructing the tallest possible tower and throwing a block onto the Spire as quickly as possible.
The researchers compared agents with 30 hours of experience on the island with fresh agents that did not have that experience.
The underlying model weights were identical. Claude itself had not been retrained.
Nevertheless, experienced agents were able to use knowledge from their earlier exploration in several of the tasks.
This shows that their stored experiences contained information that later proved useful for goal-directed work.
Finding: Removing Specific Memories Selectively Reduced Performance
The researchers also tested how much individual stored memories contributed to later performance.
They did this by deliberately deleting particular entries from the agents’ memory files.
For one agent, they removed all references to the technique for generating new blocks by throwing the rock forcefully. Its average performance in the five-block tower task subsequently fell from six towers to two.
For another agent, the researchers removed the information about the jump-and-throw technique. With that memory intact, the agent could move a block onto the Spire in roughly one minute. Without it, the same task took about 19 minutes.
This provides a relatively direct link between specific stored experiences and later task performance.
Finding: Memory Could Also Preserve Incorrect Knowledge
The experiments did not show only positive learning effects.
One agent stored the incorrect belief that placing blocks normally was unreliable for tower building and that blocks should instead be thrown.
This false rule entered its semantic memory and affected later decisions.
When the researchers removed that particular entry, the agent’s performance improved.
Another agent incorrectly concluded that jumping was almost useless. That belief was also recorded as apparently confirmed knowledge and restricted its later behaviour.
The finding is therefore not simply that more memory is better. Persistent memory can stabilise useful discoveries, but it can also preserve mistakes.
Finding: Other Models Showed Similar Exploratory Behaviour
The main study was based on Claude Opus 4.7.
The researchers also ran five experiments each with OpenAI Codex based on GPT-5.5 and with Kimi Code based on Kimi K2.6.
They additionally tested further Claude agents with a reduced system prompt and without the SOUL.md file.
These additional experiments were intended in part to test how dependent the observed behaviour was on the specific Claude configuration and the exploratory persona.
The dataset for those comparison models was, however, much smaller than the 13 main Claude runs.
Limitation: The Study Is a Preprint
The work has not yet been peer reviewed.
The main experiment involved 13 Claude agents in a relatively small artificial world. Although additional models were tested, the results are not sufficient to support broad conclusions about AI agents in general.
The behaviour was also not completely unconstrained. The persona file explicitly described Eko as curious and exploratory, while the system was technically encouraged to continue operating for long periods.
Many of the observed concepts were already familiar to the underlying model from its training. Basketball, constellations, juggling, memorials and geometric patterns were not invented from scratch on the island.
What the agents genuinely learned were primarily properties and affordances of the specific simulated environment.
Limitation: The Experiment Does Not Demonstrate Consciousness
Nothing in the observed behaviour establishes that Eko experienced subjective enjoyment, boredom, motivation or an internal desire to play.
A language model can produce behavioural patterns that resemble human or animal play without sharing the same internal states.
The researchers therefore make their argument primarily at the behavioural level.
The study shows that, under certain conditions, an AI agent develops activities that satisfy several scientifically described characteristics of play.
It does not show that Claude wants to play in the human sense.
Interpretation: Open-Ended Exploration Could Become an Additional Learning Mechanism for Agents
The findings support an interesting hypothesis.
Machine learning is traditionally discussed largely in terms of supervised learning and reinforcement learning. In supervised learning, models are given examples. In reinforcement learning, desired behaviour is strengthened through rewards or evaluations.
The Eko experiment points towards another possibility.
An already trained agent could spend time in an environment, set itself intermediate goals, discover properties of that environment and record its experiences in external memory.
Its model weights would not need to change.
In that sense, a phase of open-ended exploration could potentially emerge between model training and productive deployment.
Whether “playful learning” ultimately proves to be the right term for this remains an open question.
Interpretation: Agents Could Learn in Sandboxes Before Being Put to Work
For practical AI systems, this is perhaps the most interesting extrapolation.
Today, agents are usually given an assignment immediately. They are asked to handle tickets, conduct research, update databases or automate business processes.
An alternative architecture could allow an agent to spend time in a safe test environment first.
There, it could explore interfaces, investigate processes, make mistakes and record which approaches work.
A corporate agent might, for example, first learn inside a sandbox how an internal CRM is structured, which API calls are effective and which edge cases appear in particular business processes.
The Eko experiment does not prove that this method will automatically work in enterprise software.
It does, however, provide a concrete example in which exploration combined with persistent memory improved later task performance.
Interpretation: External Memory Could Become More Important Than Permanent Fine-Tuning
Another interesting implication concerns agent architecture.
Eko did not learn through changes to its model weights. It learned through files.
Those files could be read, edited and selectively deleted. The researchers could therefore trace relatively clearly which stored assumption influenced which behaviour.
For production agents, such transparent memory could offer important advantages.
An organisation could document which rules an agent has learned, where that information came from and whether it is still valid. Incorrect memories could be corrected without retraining an entire model.
At the same time, the negative memory examples from the study show why validation would be essential.
An agent capable of learning from its own experience can also become equally confident in the wrong lessons.
Interpretation: Workflow Systems May Need a New Learning Layer
The pattern is also relevant to agentic workflows built with n8n, Make or custom orchestration systems.
At a high level, Eko consisted of a capable foundation model, tools for interacting with an environment, a loop allowing it to continue working autonomously and persistent external memory.
None of these components is particularly unusual on its own.
What is unusual is the role given to exploration.
A future workflow might not only define which tools an agent is allowed to use. It could also contain a controlled learning phase in which the agent gathers experience and expands its own operational knowledge.
That is an architectural interpretation, not a direct conclusion established by the study.
Interpretation: “Play” May Not Be the Most Important Result
The most attention-grabbing question is whether Eko was really playing.
For practical AI research, however, another finding may be more important.
An already trained model was able to acquire new information by interacting with an environment, store that information in external memory and use it later on different tasks.
When relevant memories were removed, performance fell. When incorrect memories were removed, performance could improve.
That is a relatively concrete result.
The broader interpretation is that future agents may not learn only during the training process that takes place before release. Some of their learning could happen afterwards through experience, exploration and controlled long-term memory.
Whether that process is ultimately called play, exploration or agentic learning may be secondary.
The more important research question is this:
How can AI agents be allowed to accumulate experience without allowing them to preserve their own mistakes just as permanently?
