arXiv study finds dynamic routing outperforms fixed coordination for multi-agent systems

Enterprise multi-agent systems often default to a single coordination pattern, but a new arXiv paper by Thanh Luong Tuan suggests that dynamic selection could yield better results. The study tested 30 enterprise tasks spanning six industries, five problem classes, and four execution conditions, with three replications per cell. Four model arms were used—qwen_local, sonnet, gemma_openrouter, and an auxiliary openai cloud-validation arm—generating 1,440 outputs judged by a fixed Sonnet rubric. The goal was to determine whether coordination strategies (consensus, debate, synthesis, or single-agent) should be selected dynamically based on problem class rather than fixed globally.

The main finding is that while exact winner selection is unstable across model arms and some predicted strategies fall just short of the best observed alternative, a near-best routing claim is strongly supported. In every pre-registered model arm and problem class (including the OpenAI validation arm), the predicted strategy was within 0.10 quality-score points of the best observed condition. Notably, structured compliance verification tasks consistently favored single-agent over consensus across all arms. The study also found no reliable difference between Vietnamese-domain and English-domain tasks in strategy rankings. The key takeaway: enterprises should adopt dynamic routing as a calibrated default, not a deterministic winner-selection law.