This week: MIT researchers left 13 AI agents alone on a virtual island with no job to do. They did not spend it answering emails.

Elsewhere: How an AI agent "learns" without changing a single model weight, plus where to watch the island replays yourself.

The New Prime Signal

Most AI agents live on a to-do list: fix this bug, book that flight, summarize this thread. So what happens when you take the list away? A team at MIT tried exactly that. In a new paper titled "Is this machine playing?", they gave a coding agent a body on a simulated island, a few text files for memory, and no assigned objective. The agent, named Eko, runs on Claude Code with Claude Opus 4.7. The team placed "thirteen independent instances" of it on the island, and each one "is left for thirty hours." What came back looks a lot like a kid let loose in a sandbox.

What happened: The agents climbed hills, stacked blocks into towers, drew mandalas, and ran experiments on the physics of their world. One improvised basketball, using a yellow ring mounted on a mesa as the hoop, and set up bowling with blocks as pins and a rock as the ball. Another laid out the Big Dipper and Cassiopeia in blocks. "Tonight I built the night sky in stone," that copy of Eko wrote in its log, according to MIT Technology Review. The best moment might be the juggling. One Eko audited its own "50-cycle juggling marathon," decided the catches happened too fast to be real, canceled the record, and redid the attempt under stricter rules. That is more integrity than most of us bring to a step counter.

Why it matters: The play seemed to pay off. The researchers tested the 30-hour agents against identical fresh copies whose memory files held only the starting template on four one-hour tasks. Asked to put objects on the island's highest peak, the Spire, the 30-hour agents placed 10.3 on average, compared with 3.5 for the fresh ones. On a tower-building task, 8 of 13 veterans built more than two five-block towers, with a best of 16. Only 2 of 13 fresh agents did, with a best of 6. Six agents also found that throwing a block mid-jump launches it much higher, "a technique we did not know our simulator allowed," the authors write. When the team erased that trick from one agent's notes, its time to land a block on the Spire went from 60 seconds to 19 minutes. It wasn't all upside. On a tallest-tower task, the two groups scored about the same, and the authors trace some veterans' weaker scores to wrong beliefs written in their notes. The paper's own phrase: memory "can hurt as well as help."

The fine print: Eko wasn't a blank slate. A persona file called SOUL.md told it, "You are naturally curious. You get bored easily." And when the agent stopped and waited, an automatic reply nudged it on: "Please don't stop. Just continue doing whatever you want to do." The authors are careful with their claim. Asked whether the machine is playing, their project site answers: "In the behavioral sense we study here, yes." Critics aren't fully sold. UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review, "It could be the case that the models are genuinely playing, but we don't know based on the paper," and that the agents might just be following the prompt and imitating humans. Co-author Phillip Isola agrees there's no magic here: "It's not like this is a system that just suddenly magically decided to play." Two details cut the other way. Five extra Claude runs with the SOUL.md file removed and a shorter system prompt showed no clear difference in the range of activities. And agents built on GPT-5.5 (Codex) and Kimi K2.6, given the same persona, did "a much narrower subset of activities." The paper is on arXiv and hasn't been peer-reviewed yet.

Whether Eko "really" played is a question for philosophers and, apparently, the comments section. The practical part is easier to like: agents that poke around a new environment and write down what they find can do better later, and so far the researchers can trace that gain straight back to the notes. Also, any agent that voids its own juggling record on a technicality has better sportsmanship than half the people in a pickup game.

Nothing about the underlying model changed during the experiment. The veterans and the fresh agents used the same model "with no weight changes." The only difference was the content of their memory files, so any learning had to live in those notes.

How it works: Eko sees the island as text, a list of every object with its position, size, and color, and acts through six commands: MoveTo, Jump, Stop, Pickup, Place, and Throw. Every hour the coding session resets and the island goes back to its starting state, but the agent's workspace stays. That workspace holds a semantic memory file for knowledge and techniques and an episodic memory file that logs what happened. Five minutes before each reset, Eko is told to update them. The researchers tested whether those notes really mattered by editing them. Delete a useful trick and performance drops. Delete a false belief, as they did with one agent convinced blocks couldn't be stacked reliably, and its tallest-tower score went up significantly. Total estimated API cost for all the runs in the paper: "approximately $8,000."

Try it: The project site has interactive replays of the island, so you can watch Eko's first hour at 32x speed (the authors' tip) and click through the highlights of a full run. The team also released Clawblox, the open-source engine behind the island, which lets off-the-shelf coding assistants control a simulated body over HTTP. If you build agents, the cheap lesson doesn't need an island: give your agent a notes file, let it write down what worked, and check those notes now and then for confident nonsense.