A recent preprint on arXiv challenges a common assumption in multi-agent AI research: that building stronger individual agents automatically leads to better teamwork. The authors argue that effective collaboration depends on the communication between agents, not just on their individual capabilities.

The key insight is that communication is a two-sided problem. What one agent conveys must be aligned with how another agent interprets and uses that information. This dependency makes the communication edge—the link between two agents—a natural unit of analysis and optimization.

The paper proposes a "paired optimization" approach that targets these edges directly, rather than optimizing agents in isolation. While the abstract is brief and leaves details of the method unspecified, the framing suggests a shift from agent-centric to interaction-centric design for multi-agent systems.