Capable language models can outgrow the benefits of collaboration

Download PDF Article Open access Published: 24 July 2026 Capable language models can outgrow the benefits of collaboration Yubin Kim  ORCID: orcid.org/0000-0002-1902-38221,2, Ken Gu  ORCID: orcid.org/0000-0002-4343-15781, Chanwoo Park2, Chunjong Park  ORCID: orcid.org/0000-0003-0858-99323, Samuel Schmidgall3, A. Ali Heydari1, Yao Yan1, Zhihan Zhang  ORCID: orcid.org/0000-0001-7394-54091, Yuchen Zhuang3, Yun Liu1, Mark Malhotra1, Paul Pu Liang2, Hae Won Park2, Yuzhe Yang  ORCID: orcid.org/0000-0002-7634-82951, Xuhai Xu  ORCID: orcid.org/0000-0001-5930-38991, Yilun Du1, Shwetak Patel1, Tim Althoff1, Daniel McDuff  ORCID: orcid.org/0000-0001-7313-00821 & …Xin Liu  ORCID: orcid.org/0000-0002-9279-53861 Show authors Nature Machine Intelligence volume 8, pages 1157–1172 (2026) Cite this article Save article View saved research 1601 Accesses 2 Altmetric Metrics details Subjects Computer scienceMathematics and computing A preprint version of the article is available at arXiv. AbstractAgents, language model-based systems that can reason, plan and act with tools to accomplish tasks, are widely deployed, yet it remains unclear when multi-agent coordination outperforms a strong single agent. Here we conduct a controlled experiment that holds task prompts, tools and compute budgets constant while varying only coordination structure and model capability. Across 260 configurations spanning six benchmarks, five architectures and three LLM families, we derive a predictive model using empirical coordination metrics. Across benchmarks, single-agent baseline performance emerges as the most robust predictor of whether coordination improves or decreases performance. In particular, we identify an empirical capability-saturation threshold beyond which additional agents are unlikely to improve performance. This threshold correctly predicts the effect of multi-agent coordination on performance in 94% of validation configurations on SWE-bench Verified and Terminal-Bench. We therefore interpret this threshold as a practical selection rule rather than a universal scaling principle. A second effect, baseline-scaled error amplification, survives cluster-robust inference (Probust = 0.030) and suppor