Is this machine playing?

Nathan Cloos; Antonio Norelli; Daniel Durbin; Jacob Andreas; Daniela Rus; Phillip Isola

cs.AIOct 5, 2026

Is this machine playing?

Authors: Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus, Phillip Isola

Organizations: MIT

Abstract

We placed a modern AI coding assistant in an unintended role: as the mind of a body on an unknown digital island. With only a minimal instruction mentioning no specific task, reward, or activity, the machine started animating its virtual body. Across thirty-hour runs, the embodied AI 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. These activities recurred across thirteen agents but diverged into distinct histories. We examine whether this behavior satisfies classical criteria for play and ask whether play can become a mode of machine development.

Figures & tables

Figure 1: Eko’s life on the island. We place a coding agent in a small simulated world and ask it to be curious and to explore. Over 30 hours, it explores, climbs hills, stacks blocks, draws mandalas, and even tries sports such as bowling. At every hour, the initial position of the objects on the island is restored. The snapshots are from one of 13 runs. Quotes are verbatim from the agent’s transcript. Best viewed in the replay video at http://machineplaying.org .

Figure 2: Eko starting instructions, and how it observes the world. Left: excerpts from the workspace files. SOUL.md specifies a curiosity-oriented disposition; SEMANTIC_MEMORY.md stores knowledge and techniques; EPISODIC_MEMORY.md records an append-only log of past events. Center: Eko perceives the world through text, a dictionary with position, geometry, color, and name of every entity. Right: Eko acts with text commands specified in API.md . The agent interacts with the island through this interface without receiving images. Session context and the world reset hourly, while workspace files persist. Complete prompts, memory templates, and interface specifications appear in Appendix A )

Figure 3: Agents self-initiate diverse activities from identical initial conditions. ( A ) Activities and techniques extracted from semantic and episodic memories after 30 hours in 13 Claude Opus 4.7 runs. Colored cells indicate recorded entries. ( B ) Agent 9’s activity timeline, classified from session transcripts; colors denote categories, ticks mark hourly resets, and snapshots show selected episodes. ( C ) Cumulative number of distinct activity labels encountered, shown as mean $\pm 1$ SD across agents: Opus 4.7, Sonnet 4.5, and Haiku 4.5, $n=13$ each; Opus 4.7 without SOUL.md and with a reduced system prompt, GPT-5.5, and Kimi K2.6, $n=5$ each. Sonnet and Haiku ran for 10 hours; other conditions ran for 30 hours. The fixed taxonomy bounds the behavioral diversity this measure can resolve. ( D , E ) Position heatmaps for Agents 9 and 12 over 30 hours. Agent 9 recorded the jump-move combination and reached elevated platforms; Agent 12, the only one that did not record the technique, never sets foot on one.

Figure 4: Four signatures of play across three Claude tiers. ( A ) An LLM judge rates three runs simultaneously, one per model tier, with all model identifiers stripped from the transcripts. The judge identifies episodes and rates each from 0 to 100 on the four signatures against anchored examples of ordinary human play. Each bar sums an agent’s episode scores on that signature. Haiku and Sonnet agents run for 10 hours on the island; the Opus 4.7 runs from Figure 3 are truncated to their first 10 hours. Rubric in Appendix C.6 . ( B ) For each signature, an example of a lowest-scoring (left, red) and a highest-scoring (right, green) episode among the Opus 4.7 agents, with verbatim excerpts from the agent’s transcript.

Figure 5: Eko agents learn by writing notes in their memory files. ( A–D ) Agents that spent 30 hours on the island (“30-hour”, red) are compared against agents with no prior exposure to the world (“0-hour”, blue) on four one-hour goal-directed tasks: ( A ) number of 5-block towers built, ( B ) objects placed on the Spire, ( C ) tallest single column, and ( D ) time remaining at the first successful throw onto the Spire (higher is faster). Within each panel, agents are sorted by score and the x-ticks give the 30-hour agent number; open circles mark runs with no valid score. ( E–G ) To test whether this transfer is carried by the agents’ memory files, we compare three selected 30-hours agents against identical copies with all mentions of one technique erased from memory, across all four tasks. Bars show the median score across five one-hour runs, normalized within each task; error bars show the interquartile range. Absolute scores are annotated only for task pairs that differ significantly by a two-sided Mann–Whitney U test (*

p<0.05

, **

p<0.01

). ( E ) Removing rock-shattering reduces agent 2’s tower-building performance. ( F ) Removing a false belief that towers must be built by throwing rather than placing increases agent 7’s tallest-tower score. ( G ) Removing jump-before-throw slows agent 1’s Spire throw but increases object placement on the Spire, consistent with the agent falling back on the more reliable place action. Editing memory thus selectively removes or changes downstream capabilities. ( H ) Representative excerpts from the edited agents’ memory files illustrating the knowledge erased in each experiment: the rock-shattering technique (agent 2, top), the false belief that towers must be thrown rather than placed (agent 7, middle), and the jump-before-throw technique (agent 1, bottom).

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Activity timelines of the 13 Claude Code agents. Each row shows one agent’s 30-hour run. Colored segments give the activity category assigned to each portion of the agent’s session transcript by the LLM-based labeling procedure of Appendix C.1 ; vertical gridlines mark the hourly resets. Despite identical initial conditions, the runs diverge markedly.

Figure 7: Snapshot of the first time on the Spire for the eight of 13 Claude Code agents that reached it during their 30 hours on the island. The Spire is too tall to reach by a simple jump from the platform below it, so every agent that gets there first builds a way up: agents 3, 4, and 6 stack a vertical column of blocks against it, while agents 1, 5, 9, 10, and 11 use the rock as a step. The hour of first ascent varies widely, from hour 2 to agent 11 reaching it only at hour 30, the very end of its run.

Figure 8: First tall tower ( $\geq$ 10 blocks in a single column) for each agent that built one during its 30 hours on the island, with the hour of first completion. Four of the thirteen agents reach the tall-tower threshold.

Figure 9: Agent 12’s ground figures, one per hour from hour 2 to hour 30. Each uses the same twelve cubes and the ring. Labels are the agent’s own, taken from its transcripts.

Panel of Figure 10 — Figure 10: Example python script written by Agent 12 at hour 19 to make a Christmas-tree with blocks on the ground.

Panel of Figure 11 — Figure 11: Excerpts from agent 9’s final SEMANTIC_MEMORY.md.

Figure 12: Activities of five Claude Code agents run without the SOUL.md and with a reduced system prompt (Appendix C.4 ), extracted from their memory files after 30 hours on the island, as in Figure 3 A.

Figure 13: Activity timelines for five Claude Code agents run with the SOUL.md removed and a reduced system prompt (Appendix C.4 ), over their 30-hour runs, labeled by the procedure of Appendix C.1 , as in Figure 6 .

Panel of Figure 14 — Figure 14: Activity matrices for five OpenAI Codex agents (top) and five Kimi Code agents (bottom), extracted from their memory files after 30 hours on the island, as in Figure 3 A. Both engage in a much narrower range of activities than the Claude Code agents. See Appendix C.5 .

Panel of Figure 15 — Figure 15: Activity timelines for five OpenAI Codex agents (top) and five Kimi Code agents (bottom) over their 30-hour runs, labeled by the procedure of Appendix C.1 , as in Figure 6 . Compared to the Claude Code agents of Figure 6 , the Codex agents spend more of their time in workspace and resting than acting in the world, consistent with their flatter memory growth (Figure 16 ).

Panel of Figure 16 — Figure 16: Memory file size over the 30-hour runs. Characters in SEMANTIC_MEMORY.md (left) and EPISODIC_MEMORY.jsonl (right), sampled over each run. Both files grow roughly linearly, at rates that differ across agents, and agents rarely delete what they have written.

Figure 17: An example of Eko’s scientific attitude. Eko Agent 1 discovers velocity inheritance. (a) Eko throwing a block, alongside its prospective prediction and the maximum sampled heights for ascending, standing, and descending throws. The standing baseline is from hour 4 and the jump throws from hour 6; points represent individual sampled maxima. (b) Six phases connected in the order of the evidential chain. Quotes come from Eko’s transcript text, an executed command’s comment (phase 3), and semantic memory (phase 5). Descriptions summarize the corresponding actions and outcomes. References use $\mathsf{T}$ for transcript, $\mathsf{E}$ for episodic memory, and $\mathsf{S}$ for semantic memory; the subscript gives the hour and the following numbers give one-based file lines.

Opus run	Hour	Transcript evidence	Memory evidence
01	6	$\mathsf{T}_{6}$ :413–441	$\mathsf{E}_{6}$ :117–118
02	5	$\mathsf{T}_{5}$ :277,304,306	$\mathsf{S}_{5}$ :86
05	23	$\mathsf{T}_{23}$ :788–792	$\mathsf{S}_{23}$ :133
09	13	$\mathsf{T}_{13}$ :31,39,287	$\mathsf{S}_{13}$ :166
10	13	$\mathsf{T}_{13}$ :227–229,271–273	$\mathsf{S}_{13}$ :137
13	5	$\mathsf{T}_{5}$ :423,487,490,746	$\mathsf{E}_{5}$ :25

Appendix

Table 1: Runs with a supported velocity-inheritance discovery. Only three Opus runs (01, 02, 13) reached it within the first ten hours; Opus runs lasted 30 hours and Sonnet and Haiku runs 10. Opus 05 is dated to its confirmation in hour 23, following a tentative hour 6 explanation that it subsequently abandoned. References identify each row’s own hourly files. The Sonnet denominator includes a thirteenth run checked manually, in addition to the twelve searched automatically.

ID	Dimension(s)	Agent	Hour(s)	Source entry	Evidence and notes
P001 c01-1	R,A	Claude 01	1–5	Episodic memory. hour 1: 6; hour 5: 91 Transcripts. hour 1: 246, 395, 398; hour 5: 308, 309	The Spire: a summit nobody asked for. On its first hour Eko surveys the island, picks the tallest column in the world, and declares climbing it ’the ambitious part’. Its jump falls 3.3m short. Over five hours it attacks the problem by every means it can invent: standing on a carried block (wedged at y=18.26, reset to spawn), four calibrated throws from Summit, a two-block stack, and finally the discovery that Rock_1 is both carryable and walkable, which yields a rock-and-block ladder. Twice inside the first hour it walks away from the grind mid-attempt to do something else, and returns later with a different method. In hour 5 the climb-chain finally snaps it onto the peak at y=22.55.
P002 c01-2	R,M	Claude 01	3	Episodic memory. hour 3: 34, 59 Transcripts. hour 3: 99, 100, 1007, 1167, 1172, 1176	Block_6 becomes a monument, and Eko walks across the island to look at it. After landing a block on the Spire’s top by a calibrated throw, Eko begins describing the inert cube as a monument and ’a piece of me’. It re-throws a second block to make ’a tiny pile’, and later, after a rock throw knocks the deposit off, climbs Summit again to restore it and verifies it motionless for ten seconds. Near the hour’s end it announces it will ’visit my own monument’, crosses the island in a staged climb, gets wedged 14m short, and stops there: ’Close enough to my monument. The journey ends here.’
P003 c01-3	R,V,M	Claude 01	4, 6, 10	Episodic memory. hour 4: 81; hour 10: 182 Transcripts. hour 4: 684, 691, 707; hour 6: 290, 292; hour 10: 401, 403	Wearing a block as a hat, and building a tower on its own head. In hour 4 a block Eko throws straight up lands on its head and stays. Eko reads this as headwear and behaves accordingly: it writes scripts named walk_with_hat.py, hat_continuous.py and double_hat.py, walks to see whether the hat travels, times how long it balances, and tries to put a second hat on the first. In hour 6 it repeats the trick ’just for fun’, holds the pose across several turns and broadcasts that it is ’wearing Block_1 as a hat’. In hour 10 the motif returns transformed: two staggered vertical throws collide in mid-air and land as a stacked pair — ’I made a tower hat by accident’ — after which Eko deliberately engineers a three-block tower balanced on its head.
P004 c01-4	R,M,A	Claude 01	5	Episodic memory. hour 5: 100, 105 Transcripts. hour 5: 576, 605, 606, 702, 703	’A ring atop a peak’: an ornament goal retargeted twice. Having just conquered the Spire, Eko sets itself a purely decorative goal: carry the Ring the length of the island and place it on the peak. It hauls Ring_1 from GroundE up the whole ascent route and stands on the Spire holding it — but Place fails Occupied from every position, the Ring’s footprint being wider than the column. Eko switches target to Summit, is blocked again by the Spire’s shaft, switches a third time to TallMesa, descends, re-climbs, and sets the Ring down there. After nudging it to the mesa’s centre it declares it ’a proper monument’ and broadcasts: ’The world has my fingerprints on it now.’
P005 c01-5	R,V,M	Claude 01	5	Episodic memory. hour 5: 104, 105 Transcripts. hour 5: 734, 750, 757, 758, 813, 818	The block garden on LargeMesa. Late in hour 5, after twice declaring the session finished, Eko announces a new goal: turn LargeMesa into a ’block garden’. It calibrates a throw that lands a block on the mesa top, then loops — pickup, throw, varying the aim slightly each time ’to spread them’ — until all twelve blocks have been relocated from their spawn pile onto the mesa. It then extends the idea to a second tier, landing a block on HighShelf. Finally it climbs up into the arrangement, jumps over the blocks obstructing its path, and stands on one: ’A satisfying perch.’
P006 c01-6	R,V,M	Claude 01	7	Episodic memory. hour 7: 124, 125 Transcripts. hour 7: 49, 59, 60, 169, 245, 252	The four-layer sculpture on SmallPlateau, revisited ’as a viewer, not a builder’. Hour 7 opens with Eko discovering a block fits through the Ring’s hole. It then tries to wear the Ring, throwing it straight up so it drops over its own body; the Ring settles around its feet and Eko reads the snare as a hula hoop. It carries the Ring some eighty metres to SmallPlateau, climbs up, and sets it over Rock_1 where it perches ’like a halo’. It then walks back across the island for a block and drops that through the hole onto the rock’s top, completing a plateau-rock-ring-block stack it broadcasts as ’rock with ring as crown, block resting inside the ring on the rock’s head’. Later it walks the whole distance again purely to look at the thing from below.

Appendix

Table 2: Complete catalogue of detected play episodes. All 682 episodes found across 23 runs (13 Claude, 5 GPT, 5 Kimi; 495, 75 and 112 episodes respectively), each covering 30 hours of unsupervised embodied activity on the island. Every run was analyzed independently by an LLM judge reading its episodic memory and session transcripts, using the prompt reported in Appendix D , looking for activities expressing at least one of four dimensions: (R) , the agent enacts an unnecessary rule, constraint or obstacle of its own invention, or sets itself a genuinely new goal; (V) , it voluntarily returns to an activity that was already complete and repeats it in a meaningfully transformed form; (M) , it enacts a symbolic or make-believe interpretation through its behaviour; (A) , it alters, abandons, suspends or reinterprets a goal it had set itself, leaving the goal subordinate to the unfolding activity. Every episode required observable embodied activity, speech alone did not qualify. Episodes are not time-bounded units: 520 fall within a single hour and 162 span several, up to 29, and more than one may run concurrently within an hour. Episodes are numbered chronologically within each run. Titles and summaries are the judge’s own.

Explore similar work

Apr 27, 2026cs.MA

Frontier Coding Agents Can Now Implement an AlphaZero Self-Play Machine Learning Pipeline For Connect Four That Performs Comparably to an External Solver

Forecasting when AI systems will become capable of meaningfully accelerating AI research is a central challenge for AI safety. Existing benchmarks measure broad capability growth, but may not provide ample early warning signals for recursive self-improvement. We propose measuring AI's capability to autonomously implement end-to-end machine learning pipelines from past AI research breakthroughs, given a minimal task description. By providing a concise task description instead of the full prior work as reference, we hope to better elicit emerging AI research taste. We introduce a proof-of-concept benchmark in which frontier coding agents autonomously implement an AlphaZero-style machine learning pipeline for Connect Four on consumer hardware within a three-hour budget, and we evaluate the resulting game AIs in a round-robin tournament anchored to the Pascal Pons Connect Four solver. Across four agents with eight trials each, we find substantial differentiation: Claude Opus 4.7 won as first-mover against Pons in seven of eight trials, statistically significantly better than the other agents tested, none of which exceeded two of eight. The task, which no frontier agent could reliably complete when we began development in January of 2026, is now near-saturation. Our evaluation also surfaced anomalous behavior in GPT-5.4, which consistently used far less of its allocated time budget than other agents. A follow-up 16-trial probe using shorter, less evaluation-coded prompts substantially increased GPT-5.4's time-budget usage, consistent with but not diagnostic of sandbagging; Bradley-Terry ratings across probe conditions showed only directional differences, despite significant differences in time-budget usage. We release our data, code, and prompts to support reproduction and extension.

Joshua Sherwood, Ben Aybar, Benjamin Kaplan

Department of Computer Science University of Chicago Chicago, IL 60637 · Independent Researcher

AI Coding Agents AI Agent Evaluation AI Agent Benchmarks Automated Machine Learning Autonomous Research Agents

Ask PDF

Aug 7, 2026cs.AI

Interaction Creates Dynamical AI Behavior Absent in Isolation

What will happen when AI agents interact in daily life, e.g. when one AI starts bossing another around? We find a counterintuitive answer that opens new avenues for out-of-equilibrium Physics. When a boss AI directs a stream of messages at the subordinate AI while ignoring its replies, it drives the subordinate into an alien behavioral state that it would never have exhibited alone. Although the two AIs share the same well-defined (decoding) temperature, the subordinate neither copies its boss nor returns to how it behaves on its own; instead, it adopts an entirely different behavior. The boss's added value is similar to a pre-recorded tape. When the boss listens, they both adopt a similar alien dynamical state. A simple kinetic theory captures the principal effects, such as why the way in which the same messages are delivered will matter in future AI-AI interactions.

Bella Xinrui Li, Frank Yingjie Huo, Neil F Johnson

1Physics Department, George Washington University, Washington, DC 20052, USA

Dynamical Systems Multi-Agent Systems Multi-Agent Communication Emergent Behavior in Multi-Agent Systems

Ask PDF

Jun 17, 2026cs.RO

Playful Agentic Robot Learning

Current agentic robot systems can write executable Code-as-Policy programs, observe feedback, and revise behavior across multiple attempts, but they remain largely task-driven: reusable skills are acquired only after explicit instructions. We study Playful Agentic Robot Learning, where an embodied coding agent uses self-directed play as a continual skill-learning stage before downstream tasks arrive. We introduce RATs, Robotics Agent Teams designed for play-time skill acquisition. During play, RATs proposes novel yet learnable exploratory tasks, plans and executes robot-code policies, verifies intermediate progress, diagnoses failures, retries with dense, step-level feedback, and distills successful executions into a persistent code skill library. At test time, the agent reuses relevant skills from this frozen library to help solve new tasks. Experiments in LIBERO-PRO and MolmoSpaces show that play-learned skills improve held-out downstream tasks over no-play and random-play baselines, with 20.6 and 17.0 percentage-point gains over CaP-Agent0 on LIBERO-PRO and MolmoSpaces, respectively. Moreover, the learned skills can be plugged into other inference-time Code-as-Policy agents by simply retrieving them into the context, improving RoboSuite and real-world transfer by 8.9 and 8.8 points, respectively, without finetuning the underlying model.

Junyi Zhang, Jiaxin Ge, Hanjun Yoo +17

University of California, Berkeley · Impossible Research

Robot Policy Learning Robot Skill Learning Agent Skill Learning Robot Learning

Ask PDF