Research shows ‘more agents’ isn’t a reliable path to better enterprise AI systems

Google and MIT researchers investigated the effectiveness of multi-agent systems compared to single-agent systems. Their research challenges the assumption that more agents always lead to better AI performance. They developed a model to predict agentic system performance based on task and architectural characteristics. Adding more agents can improve certain tasks but often introduces overhead and diminishing returns. The study distinguishes between "static" and "agentic" tasks, where agentic tasks require sustained interaction and adaptability. Experiments tested various multi-agent architectures and LLM families. The research found that multi-agent systems face trade-offs related to tool use, baseline performance, and error propagation. Specific guidelines were provided regarding when multi-agent systems are beneficial. Enterprises should prioritize single-agent benchmarks and consider the sequential nature of tasks. Careful consideration of API usage and topology selection is important. Current team sizes are limited by communication overhead. Future research emphasizes the need for sparse, hierarchical, asynchronous, and capability-aware coordination.