Study Finds More AI Agents Can Make a System Dumber, Not Smarter

06:30 · July 29, 2026

Multi-agent AI systems, where several language model instances debate, critique, or divide up a task, are often marketed as a straightforward way to boost accuracy over a single model. The new study, led by Kim and colleagues, tested that assumption directly by running 260 distinct agent configurations across a range of tasks and measuring where collaboration actually paid off. The pattern that emerged was not simple: past a certain point, adding more agents introduced coordination overhead, compounding errors, and groupthink effects that made the combined system perform worse than a lone capable model working alone. The researchers describe this as language models being able to ‘outgrow’ the benefits of collaboration once individual model capability crosses a certain threshold. To make the finding actionable, the team built a predictive selector that, given a new task, recommends whether to use a single agent or a specific multi-agent architecture, and it picked the best-performing option in 87 percent of within-domain test cases. The paper argues that as frontier models keep getting individually smarter, engineering teams should stop defaulting to bigger agent swarms and instead test whether collaboration is actually earning its extra compute cost.