As developers race to build sophisticated Multi-Agent Systems (MAS), a common assumption has taken root: throwing more AI models at a problem will inevitably yield better results. However, recent research suggests this "more is better" approach might actually be sabotaging system performance.
A new analysis delving into model pool selection for multi-agent architectures—specifically examining routing, majority voting, and LLM-as-a-Judge structures—put this conventional wisdom to the test. Researchers evaluated eight different model selection strategies, analyzing how variables like model size, baseline accuracy, and response diversity impact an AI system's overall success.
The findings turn standard industry practices upside down. The study revealed that indiscriminately expanding a candidate pool with various models might increase a system's theoretical potential, but actual real-world performance often drops significantly. In many instances, a bloated, overly diverse multi-agent system performed worse than if it had simply relied on the single best model available.
So, how should AI developers optimize their agent pools? The research points to a surprisingly simple strategy: keep it in the family. The study found that selecting candidate models from within the same overarching model family consistently demonstrated the most stable and reliable performance improvements over using a standalone model.
For engineers designing the next generation of AI agents, the takeaway is clear. When it comes to Multi-Agent Systems, careful curation beats sheer volume. Strategically grouping related models will yield far more reliable results than haphazardly combining every large language model on the market.