When LLM agents collaborate, their success depends not only on their roles and information access but also on how they coordinate: who communicates with whom, what messages contain, and whether recipients can use them effectively.

This talk examines how communication topology, message format, and model interfaces shape team performance. It asks whether agents should exchange natural-language messages or learned representations of internal states, how communication can be compressed without sacrificing useful information, and whether hyperbolic geometry can efficiently encode hierarchical relationships.

Drawing on published research and controlled numerical experiments, I distinguish established findings from open questions. The goal is to touch upon how agent teams can improve their responses through better communication, not simply more messages or greater computation.