Multiplayer experiences with agents haven’t been really explored as much, and it was interesting to test the boundaries of LLMs with respect to that.
I used durable objects from Cloudflare for the backend, and a MacMini that’s always online to handle the reading and writing to iMessage.
The nice part is that it decides whether to reply in the group chat, a bit like a human would. It doesn’t always reply like you’d expect in a ChatGPT-like setting. And it can also understand images!
It was a fun project to work on, I may write a detailed technical post about it if there’s interest.
See a demo here: https://x.com/albtaiuti/status/1899445452145373201