Towards a science of scaling agent systems: When and why agent systems work

Google Research · 8 months ago · 58

Researchers evaluated 180 agent system configurations across multiple benchmarks and found that adding more agents does not universally improve performance, contrary to common industry practice. On parallelizable tasks like financial reasoning, centralized multi-agent systems improved performance by 80.9%, but on sequential tasks like planning, all multi-agent variants degraded performance by 39-70%. The study produced a predictive model with 87% accuracy in identifying optimal coordination strategies based on task properties like decomposability and tool count, enabling developers to make informed architectural choices rather than defaulting to agent proliferation.