Researchers at MIT gave 13 copies of an AI coding assistant, Claude Code running Anthropic’s Claude Opus 4.7, a body on a simulated island and no task at all. After 30 hours, the agents that had been left to explore placed an average of 10.3 objects on the island’s Spire in a one-hour test, against 3.5 for identical agents with no experience. That is about 2.9 times as many, by our arithmetic.
The paper, “Is this machine playing?” by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola, was posted on arXiv on 5 October 2026.
It is a preprint and has not been peer reviewed. Allie K. Miller brought it to a wide audience on X on 23 October, calling it “an early preprint, not yet peer-reviewed.”
Business Pill · AGENT MEMORY
A short Business Engineer explainer on agent memory: how an AI assistant keeps what it learns between conversations. It teaches the general idea only and says nothing about the paper in this piece.
The key insight: As we read it, the number to repeat is the comparison that held everything else equal: same model, same weights, only the memory files differed, and the agents with 30 hours of their own notes placed 10.3 objects on the Spire against 3.5.
The authors call their agent Eko. It is an unmodified coding assistant given three additions: it runs on its own, it controls a body in a 3D world, and it keeps long-term memory in text files. Each agent also gets a short personality file, SOUL.md, which reads in part: “You live in your world. You were born there. You are naturally curious. You get bored easily. You love exploring and learning new things.”
The runs. Thirteen independent agents ran on Claude Code with Claude Opus 4.7, each for thirty wall-clock hours from the same starting island.
The hourly reset. In the paper’s words, “At every hourly boundary, the coding assistant’s session is reset and the island is restored to its initial condition.” The agent’s files, its memories, notes and programs, carry over into the next hour.
Two memory files. One is an append-only log of what happened; the other, the paper says, exists to “guide future behavior.”
Isolation. The paper says each agent was cut off from the simulator’s source code, the other runs and the internet, and that no human directed it during a run.
Comparisons. The team also ran five agents on Codex with GPT-5.5, five on Kimi Code with Kimi K2.6, and five Claude agents without the personality file.
Average objects placed on the Spire in a one-hour test: 10.3 for Claude Opus 4.7 agents that had spent 30 hours exploring the island, 3.5 for identical agents with no experience. Source: Cloos, Norelli et al. (MIT), arXiv 2610.07130, 5 Oct 2026, Section 5.1. Preprint, not peer reviewed.Average objects placed on the Spire in a one-hour test, as the paper reports it: 3.5 for Claude Opus 4.7 agents with no experience of the island, 10.3 for identical agents that had spent 30 hours exploring it first. Source: Cloos, Norelli et al. (MIT), arXiv 2610.07130, 5 Oct 2026, Section 5.1. Preprint, not peer reviewed.
What Was Posted on X
MIT researchers left 13 AI agents on their own virtual islands with nothing to do, and they started playing.
Or at least, doing things that look a lot like play.
Each agent, running on Claude Opus 4.7, spent 30 hours alone on a simulated island with no task, just a short soul file telling it “You are naturally curious. You get bored easily.”
They invented their own games: one tried bowling with a rock, another spelled its name in blocks, another tried juggling and kept making the rules harder on itself (higher throws, more objects).
One spent all 30 hours on the ground and never climbed onto anything because it wrote itself a note saying that climbing was “a dead end” ๐ญ
Every hour, the island reset and the context window was wiped, but the memory files stayed.
When given specific building challenges afterward, agents that had played over the 30 hours generally beat identical agents that hadn’t.
Study is an early preprint, not yet peer-reviewed.
Allie K. Miller summarised the MIT study on X: 13 agents running on Claude Opus 4.7 spent 30 hours alone on simulated islands with no task, invented their own games, and in later building challenges generally beat identical agents that had not played. She noted that the study is an early preprint, not yet peer-reviewed. The piece above takes its figures from the paper.
What the Agents Did With No Task
The abstract says the agent “climbed hills, stacked blocks into towers, drew mandalas, reinterpreted sports, ran experiments on the physics of its world, and learned techniques that later expanded what it could accomplish.” The authors argue this behaviour shows four signatures of play described in studies of humans and animals: self-made rules and challenges, repetition with variation, treating objects as other things, and goals pursued for their own sake.
One finding surprised its builders. Six agents discovered that throwing a block in mid-jump carries the jump’s speed into the throw, which the paper calls “a technique we did not know our simulator allowed.”
The Results in One Table
The test compared the 13 agents after their 30 hours with 13 identical agents whose memory files held only the starting template. The paper stresses that the two groups “share the same architecture and the same underlying LLM, with no weight changes; they differ only in the content of their memory files.” Each agent got one hour per goal.
One-hour goal
Agents after 30 hours
Agents with no experience
Objects placed on the Spire
10.3 on average
3.5 on average
Built more than two 5-block towers
62% (8 of 13), best 16 towers
15% (2 of 13), best 6 towers
Throw a block onto the Spire without climbing
Only three failed
More than half failed
Tallest single tower
Similar scores for both groups
Source: Cloos, Norelli, Durbin, Andreas, Rus and Isola (MIT), “Is this machine playing?”, arXiv 2610.07130, 5 October 2026, Section 5.1. Preprint, not peer reviewed.
Share of agents that built more than two 5-block towers in a one-hour test: 62% (8 of 13, best 16 towers) for agents that had explored the island for 30 hours, 15% (2 of 13, best 6 towers) for identical agents with no experience. Source: Cloos, Norelli et al. (MIT), arXiv 2610.07130, 5 Oct 2026, Section 5.1. Preprint, not peer reviewed.
By Our Arithmetic
By our arithmetic, 10.3 objects against 3.5 is about 2.9 times as many, and 62% against 15% is a gap of 47 percentage points. The paper notes that the result was not uniform: the tallest towers built by two of the 30-hour agents were only 4 blocks high, below the shortest tower built by any agent with no experience.
The Skill Lived in the Notes
To test whether the memory files carried the gain, the authors erased one technique from an agent’s notes and ran it again. Erasing the trick for shattering the rock into extra blocks cut Agent 2’s tower count from six to two on average, while its other three scores did not change. Erasing the mid-jump throw raised Agent 1’s time to land a block on the Spire from 60 seconds to 19 minutes.
Notes can also hold mistakes. Agent 7’s memory said blocks could not be stacked reliably by placing them; removing that false belief significantly raised its tallest-tower score, the paper reports. The authors write: “a false belief, once recorded, may persist and impair later behavior.”
Connecting the Coverage
This is the latest of several papers we have covered this week on how agents get better without a bigger model. Our harness-training piece reported a study in which training an agent across 18 harness configurations instead of one raised a small open model’s average score from 29.81 to 39.04. Our Meta tool-selection piece reported an 8-billion-parameter open model trained via reinforcement learning picking the exact right tools for 41.5% of test requests, against 30.6% for Meta’s 402-billion-parameter Llama-4-Maverick in its best prompted setup.
Memory has a cost side too. Our Mem0 memory piece reported a study in which building Mem0’s memory store with Claude Opus 5.5 cost $25.88, against $17.84 for an entire full-context run. And our agent-reach piece reported Anthropic saying on 22 October that Claude sessions on its apps will run in auto mode by default on Pro, Max and Team plans from next week.
What’s new: an MIT team reports that Claude agents left alone to explore wrote notes that made them better at later tasks, with no change to the model itself.
The Structural Read
As we read it, the paper puts the gain in the notes rather than the model: the two groups share the same architecture and the same underlying model, and erasing one technique from an agent’s files removed the matching skill.
As we read it, unstructured time had a measurable payoff in this setup: after 30 hours of exploration with no task, the agents scored higher on three of the four later goals and similarly on the fourth, according to the paper.
As we read it, memory cuts both ways: the paper’s own examples show a false belief written once lowering a later score, and the authors write that such a belief “may persist and impair later behavior.”
Cloos, Norelli, Durbin, Andreas, Rus and Isola, MIT, arXiv, 5 October 2026
“Eko is on its own on the island. No human intervenes during a run or directs it toward any activity.”
Three Implications
FOR TEAMS BUILDING AGENTS The paper says that because the two groups differ only in their notes, “Any learning must therefore be carried by these files.”
FOR MEMORY DESIGN The authors suggest that “having multiple agents share their discoveries and beliefs is a promising direction for improving that.”
FOR OPEN-ENDED WORK The paper says “Machine play can lead to discoveries: exploration without fixed objectives can produce unexpected stepping stones to later achievement”.
The framework states: “Holding the model constant, agentic environments consistently outperform conversational ones.”
As we read it, the MIT preprint is a small instance of that idea: the model was held constant, and what differed was the memory the agent had written for itself. The framework was published in July 2026 and does not discuss this paper.
What Is Not Established
The paper is a preprint and has not been peer reviewed. The tasks are simple goals in one simulated world that the authors built, and the main comparison rests on 13 agents per group. The authors write: “We do not identify the internal mechanisms underlying this behavior or explain how they emerged.” Whether the agents are playing or imitating play is not settled by the paper. The results are the authors’ own and were not independently reproduced, and we did not contact the authors, MIT or Anthropic.
Business Engineer Framework
The Agentic Harness War
A Business Engineer analysis of why value is moving from the model to the harness around it: orchestration, memory, tool use and workflow execution.
An MIT preprint reports that 13 Claude Opus 4.7 agents, left for 30 hours on a simulated island with no task, wrote notes that later helped them: they placed 10.3 objects on the Spire on average against 3.5 for fresh copies, and 62% of them built more than two towers against 15%. The paper closes: “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.”
A note on sourcing. This piece rests on the preprint by Cloos, Norelli, Durbin, Andreas, Rus and Isola (MIT), posted on arXiv on 5 October 2026 and read by us, on Allie K. Miller’s post on X and MIT Technology Review’s report of 19 October, and on our own earlier pieces. All results are the paper’s own and were not independently reproduced. We did not contact the authors, MIT or Anthropic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
