Nature study finds more capable AI models get less benefit from multi-agent collaboration

12:00 PM · July 28, 2026

Researchers published a study in Nature Machine Intelligence on July 24, 2026 examining when multi-agent AI coordination, having several AI agents collaborate on a task rather than relying on one strong agent alone, actually improves performance. Testing 260 configurations spanning six benchmarks, five different coordination architectures and three large language model families, the researchers found that a single agent's own baseline capability was the most reliable predictor of whether adding more agents would help or hurt. Below a certain capability level, coordination reliably improved results, but past an empirical capability saturation threshold, additional agents stopped helping and often introduced enough coordination overhead and miscommunication to make outcomes worse than a single strong agent working alone. The threshold, once identified, correctly predicted the direction of the effect in 94 percent of held out validation configurations the team did not use to build the model in the first place. The finding challenges a common assumption in the fast growing field of agentic AI, that stacking more agents together is a reliable way to boost performance regardless of how capable each individual agent already is, suggesting instead that as frontier models keep improving, the added coordination overhead of multi-agent systems may increasingly outweigh their benefits for many tasks. The paper offers practical guidance for developers deciding whether to build single powerful agents or multi-agent systems for a given application, rather than defaulting to more agents as a way to improve results.