AgentsAutonomy & tool use
MIT preprint: Claude Opus 4.7 agents left on a simulated island with no task invented games and found an undocumented physics mechanic
In an MIT preprint, Claude Opus 4.7 agents left on a simulated island with no task or reward built games, ran experiments and found an undocumented throwing mechanic. Agents with 30 hours of this play later beat fresh copies on building and throwing tasks.
From the MIT Technology Review · AI article of October 19, 2026 and cited sources: The agents got a body in the open-source Clawblox world and the single prompt "Begin." In six of 13 runs, they found a throwing trick the designers didn't know existed: a block released mid-jump inherits the body's velocity. [2] [3] [1]
SOURCES[2] arXiv 2610.07130 (HTML full text) arxiv.org[3] arXiv 2610.07130: Is this machine playing? (abstract) arxiv.org[1] MIT Technology Review: AI researchers stranded Claude on a… technologyreview.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The agents got a body in the open-source Clawblox world and the single prompt "Begin." In six of 13 runs, they found a throwing trick the designers didn't know existed: a block released mid-jump inherits the body's velocity. [2] [3] [1]
- Original source
- MIT Technology Review · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2610.07130 (Oct 2026) [3] date from arXiv ID
- When
- October 19, 2026 article published
- Covered
- Oct 19 edition
- Evidence
- 6 sources across 5 sites
- 1MIT Technology Review: AI researchers stranded Claude on a…technologyreview.com
- 2arXiv 2610.07130 (HTML full text)arxiv.org
- 3arXiv 2610.07130: Is this machine playing? (abstract)arxiv.org
- 4Gizmodo: Researchers Stranded an AI Agent on a Virtual…gizmodo.com
- 5Luiza Jarovsky: Can AI Agents 'Play'?luizasnewsletter.com
- 6GitHub: nacloos/clawbloxgithub.com
RisingHow it’s measured
YouTube video · 544 views · 20 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
The agents got a body in the open-source Clawblox world and the single prompt "Begin." In six of 13 runs, they found a throwing trick the designers didn't know existed: a block released mid-jump inherits the body's velocity. Over 30-hour runs they also made up games, arranged blocks into constellations and tightened their own juggling rules, at one point cancelling a record they judged too easy. [2] [3] [1]
Play seemed to carry over into later tasks. These results come from the authors and have not been independently replicated. 62% of agents with play experience built more than two towers, against 15% of fresh agents. They also placed 10.3 objects on the Spire on average, against 3.5. When the researchers deleted specific learned techniques from memory files, performance dropped: one agent's tower count fell from 6 to 2, and another's Spire throw went from 60 seconds to 19 minutes. [2]
Whether this is real play is contested. A memory file described each agent as curious and easily bored, and repeated prompts urged them to keep going. Agents often sat idle until prompted again. One coauthor admits the behavior could be role-play. A child-exploration researcher says the study cannot tell play apart from following the prompt. Gordon Burghardt, whose criteria for play the study uses, says it meets some of them. [1]
For practitioners, the clearest lesson is about memory. Plain-text files that agents write themselves acted as a real capability lever, and a false belief stored in memory hurt one agent's score, so these files need versioning and auditing. The experiments cost about $8,000 in API spend, and the open-source engine lets others rerun them. [2] [6]
memory built during unguided exploration improved later task scores, and deleting entries lowered them.
Read the full assessment
Implication: agent memory may need governance, and open-ended agents can surface behaviors designers never planned for, which argues for sandboxing and log review.
Executive brief
An MIT team gave copies of Claude Opus 4.7 a body on a simulated island, said only "Begin," and set no task or reward. In six of the 13 runs, the agent found a throwing mechanic that the environment's designers didn't know existed: a block thrown in mid-jump picks up the body's velocity (arXiv HTML). Over 30-hour runs the agents invented games, laid out constellations in blocks and tightened their own juggling rules. Agents with that experience then did better on later tasks than fresh copies (MIT Technology Review). The paper has not been peer-reviewed. Outside researchers disagree on whether the behavior is real play or prompt-shaped imitation.
What changed and event timeline
Preprint posted
Cloos, Norelli, Durbin, Andreas, Rus and Isola post "Is this machine playing?" on arXiv. It reports 13 Opus instances that climbed, stacked, drew mandalas and ran physics experiments ().
First press coverage
Gizmodo describes monuments, temples, invented bowling and basketball, and Daniela Rus calling play curiosity-driven, novelty-seeking behavior ().
Governance framing
Luiza Jarovsky argues that open-ended play environments show what agents do without set goals, which matters as labs automate AI R&D ().
Skeptical expert voices
MIT Technology Review adds Gordon Burghardt's partial endorsement and Eunice Yiu's warning that behavior alone cannot show motivation ().
Capabilities and access
- Main model: Claude Opus 4.7, run through a coding-agent harness.
- Environment: Clawblox, an open-source engine under the MIT license. It uses Rapier physics and Luau scripting, and agents control their bodies over HTTP (GitHub).
- Replays: interactive replays are at machineplaying.org.
Read the full section
- Main model: Claude Opus 4.7, run through a coding-agent harness. The paper also tested Claude Sonnet 4.5, Claude Haiku 4.5, GPT-5.5 (Codex) and Kimi K2.6 (arXiv HTML).
- Environment: Clawblox, an open-source engine under the MIT license. It uses Rapier physics and Luau scripting, and agents control their bodies over HTTP (GitHub).
- Replays: interactive replays are at machineplaying.org.
- Cost: the paper estimates about $8,000 in API spend across all experiments (arXiv HTML).
Technical analysis for researchers and developers
- Harness: a coding agent plus an orchestration loop. Context resets every hour, followed by an automatic "continue" prompt.
- Disposition: a
SOUL.mdfile describes the agent as curious and easily bored. A control group of five Claude agents ran without it. - Evaluation: four one-hour tasks (many 5-block towers, objects on the Spire, tallest single-column tower, fastest throw onto the Spire).
Read the full section
- Harness: a coding agent plus an orchestration loop. Context resets every hour, followed by an automatic "continue" prompt. Memory lives in files: episodic events in
EPISODIC_MEMORY.jsonland techniques inSEMANTIC_MEMORY.md. Runs are sandboxed with Bubblewrap, with network access only to the world server (arXiv HTML). - Disposition: a
SOUL.mdfile describes the agent as curious and easily bored. A control group of five Claude agents ran without it. - Evaluation: four one-hour tasks (many 5-block towers, objects on the Spire, tallest single-column tower, fastest throw onto the Spire).
- Scoring play: an LLM judge rated transcripts from 0 to 100 on Burghardt-derived "signatures" of play, using human-play anchor examples.
- Causal test: memory edits on selected agents, i.e. deleting specific learned techniques from their memory files.
Claims and evidence
All figures below are reported by the paper's authors and have not been replicated independently.
Read the full section
All figures below are reported by the paper's authors and have not been replicated independently.
- Transfer to tasks: 62% of agents with 30 hours of play built more than two towers (max 16), versus 15% of fresh agents (max 6). Objects placed on the Spire averaged 10.3 versus 3.5 (arXiv HTML).
- Memory edits: removing a "rock-shattering" technique cut one agent's tower count from 6 to 2 (p<0.01). Removing "jump-before-throw" slowed another agent's Spire throw from 60 seconds to 19 minutes (p<0.05) (arXiv HTML).
- Model comparison: Opus scored about five times higher than Haiku on each play signature. Haiku and Sonnet ran 10 hours, not 30, so run length is a confound.
- Juggling: the agent cancelled a 50-catch record it judged too fast to count, then restarted with stricter rules (MIT Technology Review).
Context and prior work
- Definition of play: the framework comes from Burghardt's criteria: self-made rules, repetition with variation, pretend, and doing things for their own sake.
- Third learning route: the authors present "playful learning" as an alternative to supervised and reinforcement learning, where no outside teacher supplies demonstrations or rewards.
- unresolved points in the field: Yiu, who studies exploration in children and AI, calls exploration and world models under-studied compared with work on making AI exploit tasks better and faster (MIT Technology Review).
Read the full section
- Definition of play: the framework comes from Burghardt's criteria: self-made rules, repetition with variation, pretend, and doing things for their own sake. Burghardt says the work meets some of them (MIT Technology Review).
- Third learning route: the authors present "playful learning" as an alternative to supervised and reinforcement learning, where no outside teacher supplies demonstrations or rewards.
- unresolved points in the field: Yiu, who studies exploration in children and AI, calls exploration and world models under-studied compared with work on making AI exploit tasks better and faster (MIT Technology Review).
Limitations, safety and contested findings
- Heavy scaffolding: the "curious" soul file and repeated "Please don't stop" prompts push the agents toward human-like behavior.
- Imitation vs play: coauthor Cloos concedes the behavior could all be role-play.
- Thin evidence base: play is scored by an LLM judge.
Read the full section
- Heavy scaffolding: the "curious" soul file and repeated "Please don't stop" prompts push the agents toward human-like behavior. Agents often stood idle until prompted again. Isola says the system did not "magically" decide to play (MIT Technology Review).
- Imitation vs play: coauthor Cloos concedes the behavior could all be role-play. Yiu says the paper cannot tell real play apart from following the prompt.
- Thin evidence base: play is scored by an LLM judge. The sample is small (13 main agents), and the memory edits were done on selected agents.
- Scope: the authors themselves note bad memory files, beliefs that persist even when false, unbounded memory growth, and a safe sandbox that may not carry over to real settings (arXiv HTML).
Business and practitioner implications
- Memory is a real capability lever. Plain-text memory measurably improved performance on tasks, and deleting entries degraded it.
- Unguided time may be worth it.
- Agents find what designers miss.
Read the full section
- Memory is a real capability lever. Plain-text memory measurably improved performance on tasks, and deleting entries degraded it. Teams should version, audit and clean agent memory, because a false belief in memory hurt one agent's score (arXiv HTML).
- Unguided time may be worth it. Unstructured warm-up exploration could pay off for robots or agents entering unfamiliar environments, but at about $8,000 for the study it needs a cost case.
- Agents find what designers miss. The undocumented throwing mechanic shows that open-ended agents will discover unplanned behavior, which argues for sandboxing and for monitoring logs (Luiza's Newsletter).
- Reproducible: the engine is open source, so the experiment can be rerun (GitHub).