I've been doing my new projects with two AI agents: one to write the program, one to build a server to host it. And I've been relaying commands back and forth between the two. I just realized I could have one Claude Code session manage the other for me. I'm delighted to find it knows all about doing that and has a messaging system for the two to talk to each other.

I spent my graduate career in the 1990s working on multi-agent systems. I never imagined anything like this actually working. The language capabilities of LLM have unlocked so many possibilities.