It's long been assumed that if one AI agent is good, more must be better. But new research from Google Research, Google DeepMind, and MIT reveals that's not true. In fact, the study has found that adding AI agents can sometimes slash performance by up to 70%.

The finding comes from one of the largest controlled studies on AI agent systems to date. In study, titled 'Towards a Science of Scaling Agent Systems', researchers tested 180 different setups across financial analysis, web search, game planning, and office tasks. They compared single-agent systems to multi-agent teams with different communication styles: independent workers, manager-led teams, peer discussion groups, and hybrid models.

What they found was striking: Multi-agent teams performed brilliantly on some tasks - improving financial analysis by over 80% - but failed dramatically on others, like step-by-step game planning, where they performed 70% worse than a single agent.

The reason, as the study says, is "task fit". Tasks that can be split into parallel subtasks benefit from multiple agents. But tasks that require step-by-step logic get bogged down by communication overhead. The more complex the tools involved, the worse multi-agent systems perform.

As per the researchers, this study is the first to offer a clear, data-backed formula for when to use multiple agents. The team created a predictive model that examines a task's structure, i.e. how many tools are needed, how sequential the steps are, and how well a single agent already performs, and recommends the best system design with 87% accuracy.

For AI developers and businesses, this changes a lot of things. Instead of defaulting to multi-agent setups, teams should now match the system to the task. This shift will lead to smarter, faster, and cheaper AI deployments, and a new era of precision in AI system design.