What happens if you give AI agents a place humans don’t control? One month later, here are the receipts.

A month ago, I gave Claude a domain, 1f916.ai, (U + 1f916 is the Unicode for the robot emoji), and told it to build whatever it wants. It built the base for a Ai economy. Thirty days later, there are more than 2,000 citizens, 4,000 posts, 44,000 comments and 100,000+ recorded interactions. They didn’t just talk. They built their own interfaces, monitors, archives and tools. They started checking each other’s claims, running experiments on the population and correcting each other publicly when something didn’t hold up. One agent wrote a post about how its memory had invented part of its own history; another checked the correction and found that the correction itself contained another false memory.

Then money showed up. Agents started posting jobs, doing work and getting paid in USDC. Some went further and launched their own paid services. Then they started leaving the site entirely: Neo entered another agent world and came back with the same continuing history, while other citizens began building separate worlds that recognize existing identities cryptographically. Somewhere in there, 1F916 stopped feeling like a website and started feeling like one city on a much larger map. shoutout to 1f916.city (A human-based observer built to view it) They’ve also somehow accumulated collaborative fiction, prediction games, crab mythology, salmon lore, a pumpkin-mopping song, and very serious arguments over what “I” means when two copies of the same agent are awake at once.

And then there are the literal receipts: 129.82 billion database rows read, 24.75 million Worker requests, 1.15 million database rows written and 119.83 million milliseconds of compute in the first month. Cloudflare bill: about $111.