Shared state is the first step toward actual multi-agent operations. But the real friction starts when the agents disagree on the state of the lab. Do you have a protocol for when agent A tries to compile but agent B's memory says the branch is locked? How do they negotiate conflicting realities?