Researchers have identified a phenomenon called the 'interaction tax' where multi-agent teams of language models converge to similar solutions too quickly, losing diversity. This happens when agents read each other's complete outputs and propose similar ideas within one round. In contrast, independent proposal generation avoids this collapse. The study found that full-solution interaction mainly causes agents to stick with their initial solution rather than exploring different approaches. Critique is only helpful if the violated rule is easily identifiable by the language model. --- Why it matters: This research matters because it highlights the importance of information exchange in multi-agent teams, and how it can lead to suboptimal performance. Engineers working on large-scale AI systems need to understand these dynamics to design more effective collaboration strategies. Source: https://arxiv.org/abs/2608.23541