What the agents did for thirty hours on the island
On 5 October 2026, Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola posted “Is this machine playing?” to arXiv. Thirteen independent agents, each running Claude Code with Claude Opus 4.7 on an embodied architecture the authors call Eko, spent thirty wall-clock hours on a simulated island. Their instruction named no task, reward or activity, and no human intervened during a run. Sessions and the world reset every hour, while each agent’s workspace, including the notes and memory files it wrote, persisted.
The agents climbed hills, stacked blocks, drew mandalas, invented variations on sports and ran experiments on the physics of their world. Those activities recurred across agents and diverged into distinct histories. One example the authors highlight is a throw that reaches higher when the agent releases the block mid-jump, an affordance the paper says the simulator’s designers did not know they had built. The paper reports seven agents describing that technique; TechCrunch’s account says six learned it.
The paper asks whether this behaviour satisfies classical criteria for play, and answers yes in the behavioural sense while leaving phenomenal play open. This article calls it play-like behaviour for that reason.
The model is named in the paper body, not only in the coverage. Five other agents ran on Codex with GPT-5.5, five on Kimi Code with Kimi K2.6, and five Claude agents ran with the soul file removed. MIT Technology Review describes the island as “loosely based on the video game Roblox”. In the paper it is an instance of Clawblox, an engine the authors released under an MIT licence, written in Rust with Rapier physics and with worlds scripted in Luau, the language Roblox uses.
The four criteria the authors applied, and who scored them
The paper defines four signatures of play: rule-making, varied recurrence (repetition without stereotyping), make-believe (treating an object as something other than its plain use) and autotelicity (pursuing something for its own sake). The authors assembled them from Huizinga, Caillois, Suits, Burghardt and others. Burghardt’s own framework has five diagnostic criteria, and the paper notes that repetition without stereotypy is one of the five for him. These are the four criteria the authors applied, rather than a standard inherited intact.
The scoring was theirs as well. One LLM judge received transcripts and episodic memory and rated episodes from 0 to 100 against anchored examples of ordinary human play. Model identifiers were stripped from the transcripts, making the judge blind to which model tier a run came from. No human inter-rater reliability is reported. In that comparison Opus scored roughly five times Haiku on every signature, with Sonnet in between.
The transfer test: large gains on two goals, none on the tallest tower
Thirty-hour agents and control agents that had run for zero hours were compared on four one-hour goals. On the Spire, the experienced agents placed an average of 10.3 objects against 3.5 for the controls. On the many-5-towers goal, 62% of them (8 of 13) built more than two towers against 15% (2 of 13). A “2.9x more objects” figure circulates in an aggregator headline; 2.94 is the ratio between those two means, and the paper does not state it. The best control run on the object goal reached 12 against the experienced group’s 16, per the authors’ project site, so the ranges overlap. Each agent was evaluated once for one hour.
The tallest-tower goal shows no advantage at all: the site gives a mean of 11.2 blocks for the thirty-hour agents against 12.1 for the controls, and two of the thirty-hour agents built four-block towers, shorter than the shortest control tower. On the throwing goal, more than half the control agents failed while three of the thirteen experienced agents did. The benefit of the thirty hours depends on the goal.
Removing one memory line changed what an agent could build
The causal test erases a single technique from an agent’s memory and compares it with an intact copy over five one-hour runs per task, using a two-sided Mann-Whitney U test. Removing a rock-shattering technique from one agent cut the number of five-block towers it built from six to two on average, leaving the other goals unchanged. The world starts with twelve blocks, so towers beyond the initial two depend on breaking the rock. Removing a false belief that towers had to be thrown rather than placed improved a second agent’s tallest-tower score, and removing a jump-before-throw technique slowed a third agent’s throws at the Spire while improving its object placement.
The precise numbers sit in figures and appendices that could not be read in full. Two third-party AI summaries give conflicting timings for the jump-before-throw effect, so neither belongs in print without the figure itself.
The setup that encouraged the behaviour
Every agent’s workspace began with a SOUL.md file describing its character. The excerpt in the paper reads: “You live in your world. You were born there. You are naturally curious. You get bored easily. You love exploring and learning new things.” A comparison group of five Claude agents ran without SOUL.md and with a reduced system prompt.
An orchestrator also kept the agents moving. It starts each session with “Begin” and automatically prompts the assistant to continue whenever it stops, with no task or success condition attached. MIT Technology Review reports the continuation prompt read, “Please don’t stop. Just continue doing whatever you want to do.” The paper confirms the automatic loop; that exact sentence appears only in the coverage, and the appendix that would settle it was not readable.
The authors’ own framing is candid about this. Phillip Isola told MIT Technology Review: “It’s not like this is a system that just suddenly magically decided to play … It’s an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us.”
What the paper cannot show about play
The authors apply their criteria while “leaving open the phenomenological question of whether it is experiencing enjoyment, absorption, or fun”, and state that they “do not identify the internal mechanisms underlying this behavior”. Eunice Yiu, who studies exploration in children and AI at UC Berkeley, told MIT Technology Review: “It could be the case that the models are genuinely playing, but we don’t know based on the paper.” She added that from behaviour alone “it could be the case that they’re just trying to abide by the prompt, trying to imitate humans.”
Gordon Burghardt’s view arrives as a paraphrase rather than a quotation: MIT Technology Review reports that he agrees the work meets some of the criteria for play he identified in living creatures. No direct comment from him appears in the coverage.
The paper is an arXiv preprint posted on 5 October 2026 and has not been peer reviewed. No independent replication or published critique turned up, and that is not proof there is none. Appendices covering the exact continuation prompt, the judging rubric and the ablation figures, along with the paper’s limitations section, could not be read in full. The code and a 26.6 GB dataset of replays, transcripts and workspaces are public.
Why the learning claim still matters
The interesting claim is about how agents acquire skills. Antonio Norelli, one of the authors, told MIT Technology Review: “There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards.” Nathan Cloos put it this way: “All of this could be role-playing … But we show that it’s not just hallucinating the play. This play is also contributing to learning.”
Two results support that reading. Techniques found during unsupervised time carried over to tasks the agents were tested on later, and deleting one technique from memory degraded the specific goal that depended on it. Both were measured in a small, fully observable, text-based world with hourly resets and thirteen agents per group, so the transfer is not evidence that this regime beats task-based training in general.
The authors’ closing suggestion is about how to build agents, not about what the agents felt: “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.”
