Google researchers Yubin Kim and Xin Liu conducted a large-scale evaluation of 180 agent configurations to uncover scaling principles for AI agent systems. Their findings challenge the common belief that simply adding more agents improves performance in all scenarios. They found that while multi-agent coordination can enhance performance in parallelizable tasks, it often degrades effectiveness in sequential tasks. For example, in financial reasoning, centralized coordination improved performance by 80.9% over a single agent, but in sequential tasks like planning, multi-agent setups resulted in a performance drop of 39-70%. The researchers defined "agentic" tasks as those requiring sustained interactions, iterative information gathering, and adaptive strategy refinement. They assessed five architectures: a single-agent system (SAS) and four multi-agent variants (independent, centralized, decentralized, and hybrid) across benchmarks like Finance-Agent and PlanCraft. Notably, independent multi-agent systems significantly amplified errors, while centralized systems limited error propagation. The authors also developed a predictive model that accurately identifies the best architecture for 87% of unseen tasks based on measurable properties like tool count and task decomposability. The research shifts the focus from heuristics, like the idea that "more agents are better," to a data-driven understanding of when and how to deploy multi-agent systems effectively. This predictive capability allows developers to make informed decisions based on task characteristics rather than relying on assumptions. As foundational models like Google Gemini evolve, the study emphasizes that architectural choices will play a critical role in maximizing the potential of AI agents.