Agent systems, particularly those based on language models, are increasingly central to AI applications. However, the factors influencing their performance are not thoroughly understood. The authors aim to fill this gap by establishing quantitative scaling principles for these systems. They define agentic evaluation and explore how aspects like the number of agents, their coordination structures, model capabilities, and task properties interact to affect performance. The study examines four benchmarks: Finance-Agent, BrowseComp-Plus, PlanCraft, and Workbench, using five canonical agent architectures, which include both single-agent and various multi-agent systems. They analyze 180 configurations across three families of language models. Key findings reveal a trade-off between tool coordination and task complexity, with tool-heavy tasks suffering more from the overhead of multiple agents. Performance diminishes beyond a certain point when single-agent capabilities reach around 45%. Error amplification also varies by coordination type; independent agents significantly magnify errors, while centralized systems do a better job of containing them. Centralized coordination boosts performance by over 80% on tasks that can be executed in parallel. In contrast, decentralized coordination proves advantageous for web navigation, improving performance by 9.2% compared to a mere 0.2% increase for centralized systems. However, for tasks requiring sequential reasoning, all multi-agent configurations performed worse, with declines ranging from 39% to 70%. The framework developed can predict optimal coordination strategies for 87% of held-out configurations, and validation with a newer model, GPT-5.2, confirmed that four out of five scaling principles apply to unseen models.