最权威AI Agent避坑指南来了，智能体越多死得越快，效率最高暴跌70%

Core Insights - The recent paper by Google DeepMind and Google Research challenges the prevailing belief in the AI community that "more agents are better" [3][5] - The research indicates that blindly increasing the number of agents is not only costly but also ineffective, with a critical conclusion that 3-4 agents represent the optimal number for current technology [3][35] Findings on Agent Systems - The "scale paradox" suggests that as task complexity increases, having more agents can lead to quicker failures, with 3-4 agents being the "golden ratio" [3][6] - There is diminishing marginal returns for agents; if a single agent achieves over 45% accuracy, adding more agents can result in negative returns [8][10] - The effectiveness of multi-agent systems is contingent on task characteristics, emphasizing the importance of matching architecture with task attributes rather than merely increasing agent numbers [4][14] Three Fundamental Laws Governing Agents - More tools lead to higher chances of "crashing" in multi-agent systems due to increased communication costs, especially when tasks require more than 16 tools [6][7] - Stronger individual agents reduce the utility of adding more agents, as communication and alignment costs outweigh benefits [8][9] - Different collaborative architectures have varying error amplification effects, with independent multi-agent models amplifying errors by 17.2 times compared to centralized models, which control errors to 4.4 times [11][12] Task and Architecture Compatibility - Multi-agent systems are not universally beneficial; their performance is highly dependent on the compatibility of architecture with task requirements [13][14] - Tasks can be categorized into three types based on their interaction with multi-agent systems: - Tasks that can be decomposed benefit from multi-agent collaboration, showing performance improvements of up to 80.9% [15] - Tasks with strict sequential dependencies suffer performance declines of 39% to 70% when using multi-agent systems [16][18] - Tasks that require both exploration and execution show mixed results, with performance varying significantly based on architecture design [19][21] Economic Analysis of Multi-Agent Systems - Multi-agent systems exhibit a drastic drop in efficiency, with centralized architectures achieving only 21.5 successful outcomes per 1000 tokens compared to 67.7 for single agents [30] - The number of communication rounds increases quadratically with the number of agents, leading to shallow reasoning and declining performance [31][34] - The optimal number of agents is identified as 3-4, beyond which communication costs dominate and lead to negative marginal returns [35]