A new study finds that piling more AI models into a multi-agent system often makes it perform worse, not better.
Researchers evaluated eight strategies for selecting which models to combine in these systems, testing criteria like model size, accuracy, and how much variety exists between candidates' answers. They ran the comparison across two common multi-agent designs: routing, which picks a model before it generates an answer, and after-the-fact methods like majority voting or using one model to judge another's output. The tests covered difficult scientific benchmarks. Across the board, the researchers found a wide gap between what a model pool could theoretically achieve and what it actually delivered, and expanding the pool of candidate models frequently dragged performance below that of the single best model working alone.
That is a useful corrective for anyone assuming a bigger roster of models is automatically safer or smarter. The one selection strategy that reliably beat a standalone model was choosing multiple models from within the same family, rather than mixing models from different vendors or architectures. For teams building agent pipelines, that is a specific, testable design rule rather than a vague best practice.
The paper's real message is that model diversity is not free: bolt on the wrong mix of models and you add instability, not insurance.