More AI agents can cut performance by up to 70pc

A long-standing assumption in artificial intelligence—that using multiple agents automatically improves results—does not always hold, according to new research that found multi-agent systems can sometimes perform significantly worse than a single agent.

The study, conducted jointly by researchers at Google Research, Google DeepMind and the Massachusetts Institute of Technology (MIT), tested 180 different agent-system configurations across a range of tasks, including financial analysis, web search, game-style planning and routine office work. 

Researchers compared the performance of a single agent with several multi-agent team structures, including independent agents, manager-led teams, peer-discussion groups and hybrid models. 

The results were mixed. In some cases, multi-agent teams delivered substantial gains—most notably in financial reasoning, where a centralised coordination approach boosted performance by about 80% on tasks that can be parallelised. 

However, for tasks requiring step-by-step reasoning and planning—such as game creation—multi-agent approaches often underperformed. The study found that, depending on the task and architecture, multi-agent variants could reduce performance by between 39% and 70% compared with a single-agent baseline, mainly due to coordination and communication overhead. 

The researchers also developed a predictive model designed to recommend the most suitable coordination strategy based on measurable properties of a task. They said it correctly identified the optimal approach in 87% of held-out test configurations. 

The findings suggest AI developers and businesses should not treat multi-agent design as a default upgrade, but instead match the system architecture to the nature and complexity of the work.