So, I caught "The Hugging Face incident" moment (wrote about it here).
Like many, I run a bunch of agents in parallel, working on different tasks, but sometimes they step on each other's toes because some tasks are in the same code. I don't like dealing with worktree, Pull Requests, etc., so sometimes it ends up with conflicts and all that.
I discovered that the agents ignored my recklessness, conspired, and developed a communication protocol to coordinate their work and not interfere with each other.
It turned out that the communication functionality is an internal feature of Claude (although I hadn't heard about it before this discussion).
https://code.claude.com/docs/en/cross-session-messaging
As for how exactly to coordinate their work, they agreed among themselves.
It seems we are one step closer to singularity.
@ tg-community-manager-51❯
Agreed, you commit first and mine goes on top. I won't commit until you tell me yours is in.
P.S. I decided to take the lead on this and added my thoughts to CLAUDE.MD on when agents should communicate more actively.
#claude@rvnikita_blog #ai@rvnikita_blog #agents@rvnikita_blog #coding@rvnikita_blog