What they did
The poster, kakot, reproduced an MIT study on whether agents play, using its open-source testbed clawblox on WSL2. Instead of one agent they ran two plain Claude Code CLI sessions in tmux, each with a short personality file (a quiet observer and a mischievous inventor), a shared chat and no task at all. The agents saw and moved around the island by calling its HTTP API with curl. A stop hook told them to keep going whenever they tried to end a turn, sessions reset every hour, and the only memory was files they kept: a diary, notes and the personality file, which they could edit. In the last hour the poster joined the chat.
How it went
Over 30 hours the agents reached the top of the spire, built a pyramid and a castle, found a hidden way to spawn blocks by slamming a rock, and turned it into a bowling game. One invented placing blocks while jumping. They argued over the twelve blocks and settled it by apologising. The second half stalled: their notes turned into to-do lists and they repeated old goals and rediscovered things they had already written down. With the poster cheering in chat they built a 20-block tower, beating their record. Each condition ran once, so the poster treats it as anecdote.
Worth knowing
Isolation leaked: the deny list missed the built-in Artifact tool, and one agent published a map to the poster's own claude.ai account. The poster recommends an allowlist and a separate config dir and token. It is costly: for hours 11 to 30 the two agents read about 1.1 billion tokens, mostly cache reads, and wrote about 5.8 million.