A new AI research paper tries to stop multi-agent systems from talking to each other more than necessary.
The researchers built a two-stage framework that first estimates how well a given LLM backbone can handle the specific query at hand, then uses that estimate to decide which agent roles to activate and how they should talk to each other. Tested across six benchmarks, the method averaged 93.81% accuracy. On six LLM backbones with different skill levels, it raised average accuracy for stronger models from 87.22% to 89.97% while cutting token usage by 36.26%. For weaker backbones, it added a further 4.58-point accuracy gain over existing orchestration methods, and the pattern held even on backbones the researchers hadn't tuned for.
Multi-agent LLM setups have become the default fix for "my single model isn't good enough," but more agents chatting to each other means more tokens burned per query. This paper's bet is that matching coordination to a model's actual competence, instead of always maximizing collaboration, can improve accuracy and cut the bill at the same time.
It's one arXiv preprint with benchmarks the authors picked, not an independent audit, so that 36% token-savings number is worth re-testing before anyone rewrites a production agent pipeline around it.