Research Examines How Multiple AI Agents Impact Performance

A new study conducted by researchers from Google Research, Google DeepMind, and the Massachusetts Institute of Technology explored the limits and advantages of multi-AI agent systems. Published in the journal Nature Machine Intelligence, the findings showed that adding more AI agents to a system can sometimes degrade performance.

Scope of the Research

Computer scientists worldwide are developing a wide variety of AI systems based on single AI agents or multiple interactive agents that exchange information and collaborate.

Researchers from Google Research, Google DeepMind, and the Massachusetts Institute of Technology conducted a controlled experiment to examine the potential advantages and limitations of multi-AI agent systems.

Experiment Configuration

Led by Yubin Kim, the team evaluated a single-agent system alongside four canonical multi-agent variants, using six different agent benchmarks and models from three large language model families.

Across 260 experimental configurations, the researchers analyzed how agent organization affects performance while keeping task instructions, available tools, and the computational budget comparable.

Predictive Model and Baseline

The study revealed that how well a single agent already performs is highly informative when deciding whether or not to add more agents.

The researchers developed a predictive model that successfully selected the top-performing AI architecture in 87 percent of the cases.

The Role of Task Structure

In a financial analysis benchmark, the best multi-agent system improved performance by approximately 81 percent by examining independent pieces of information in parallel.

In contrast, on a sequential planning benchmark, multi-agent systems significantly degraded performance by up to 70 percent as coordination costs and error propagation outweighed the benefits of collaboration.