The Interaction Tax: Communication Erases Diversity in Multi-Agent Teams

In this work, we further study the true reasons behind this behavior, by looking deeper into how and when agent interactions help. We find that when different agents read each other's work, they reach a consensus immediately, erasing the diversity that motivated using different models. We call this the interaction tax. This interaction tax depends on the task structure. For open-ended tasks, the first round of interaction makes solutions worse 57% of the time, because agents converge to one final answer instead of exploring different approaches. Letting agents independently work on solutions first can mitigate this effect. On tasks with clear rules, e.g. Knapsack, critique helps agents fix specific errors and raises success rate from 0% to 47-73% (p<0.003). These findings suggest a design principle for multi-agent systems: interaction should be used selectively, preserving independent exploration for open-ended problems and applying critique only when errors are concrete and verifiable.