Agent Scaling Laws

Definition

Empirical laws governing how multi-agent system performance changes with the number of agents, tools, and task characteristics. Derived from 180 controlled configurations across 5 architectures, 3 model families, and 4 benchmarks (Google Research + MIT).

描述多智能体系统性能如何随智能体数量、工具和任务特征变化的经验规律。基于涵盖5种架构、3个模型系列和4个基准测试的180种受控配置得出（Google Research + MIT）。

Details

Three Scaling Laws

Law 1: Tool-Coordination Trade-off More tools in the environment makes multi-agent coordination disproportionately costly (beta = -0.267, p < 0.001). Token budgets fragment across agents.

**定律1：工具-协调权衡** 环境中的工具越多，多智能体协调的成本便不成比例地增加（beta = -0.267, p < 0.001）。令牌预算在智能体之间呈碎片化分布。

Law 2: Capability Ceiling When single-agent accuracy exceeds ~45%, adding more agents hurts rather than helps (beta = -0.404, p < 0.001). The coordination overhead outweighs any parallelization benefit.

**定律2：能力上限** 当单体智能体的准确率超过约45%时，增加更多智能体不仅无益，反而有害（beta = -0.404，p < 0.001）。此时，协调开销超过了并行化带来的收益。

Law 3: Topology-Dependent Error Amplification

• Independent (no communication): errors amplify 17.2x
• Centralized (orchestrator): errors amplify 4.4x
• The communication topology determines how badly errors propagate.

**定律 3：依赖拓扑的误差放大** - 独立模式（无通信）：误差放大 **17.2 倍** - 集中式（协调器）：误差放大 **4.4 倍** - 通信拓扑决定了误差传播的严重程度。

Predictive Model

A mixed-effects model achieves R² = 0.524 and 87% accuracy in recommending the right architecture for new tasks. Validated on GPT-5.2 (released after the paper).

混合效应模型在新任务的架构推荐中实现了 R² = 0.524 和 **87% 的准确率**。该结果在 GPT-5.2（于本文发表后发布）上得到了验证。

Limitations

• Assumes fixed token budgets
• Synchronous collaboration only
• Homogeneous agents (same model for all)

- 假设固定的令牌预算 - 仅支持同步协作 - 同质智能体（全员使用同一模型）

Connections

• Related to: Multi-Agent Architectures
• Mentioned in: Agent Scaling Laws Paper

- 相关条目：[[ai-agent-architecture/concepts/multi-agent-architectures|多智能体架构]] - 提及于：[[ai-agent-architecture/sources/agent-scaling-laws-paper|智能体扩展定律论文]]