I’m trying to understand the AWS Multi-Agent Collaboration (MAC) feature and how it differs from modelling agents as LLM Tools.

Intuitively, it feels like MAC is just an extension of the LLM Tools approach, where each tool could also make an LLM call. However, AWS has chosen not to model MAC as "agents can be LLM tools as well." I realize I’m missing some key insights here.

Could someone provide scenarios that illustrate what goes wrong or becomes unachievable if we think of MAC as merely modeling agents as tools? This would really help me grasp the distinction.