Multi-Agent Architectures
Definition
Multi-agent architectures are topology patterns that define how multiple AI agents communicate, coordinate, and share information to solve tasks. Research identifies five fundamental patterns, each with distinct trade-offs.
多智能体架构是定义多个AI智能体如何通信、协调及共享信息以解决任务的拓扑模式。研究识别出五种基本模式，每种模式都具有不同的权衡。
Details
Five Architecture Patterns
| Pattern | Description | Best For | Risk |
|---|---|---|---|
| Single-Agent (SAS) | One agent handles everything | Sequential reasoning, simple tasks | Context overload |
| Independent | Multiple agents, no communication | Embarrassingly parallel tasks | Error amplification 17.2x |
| Centralized | One orchestrator delegates to workers | Parallelizable tasks (e.g., financial analysis) | Orchestrator bottleneck |
| Decentralized | Agents communicate peer-to-peer | Dynamic exploration (e.g., web navigation) | Consensus overhead |
| Hybrid | Mix of centralized and decentralized | Complex multi-domain tasks | Design complexity |
| 模式 | 描述 | 适用场景 | 风险 | |---------|-------------|----------|------| | **单智能体 (SAS)** | 一个智能体处理所有事务 | 顺序推理、简单任务 | 上下文过载 | | **独立模式** | 多智能体，无通信 | 高度并行任务 | 错误放大17.2倍 | | **中心化模式** | 一个编排者向工作智能体分配任务 | 可并行化任务（例如，金融分析） | 编排者瓶颈 | | **去中心化模式** | 智能体之间点对点通信 | 动态探索（例如，网页导航） | 共识开销 | | **混合模式** | 中心化与去中心化的结合 | 复杂的多领域任务 | 设计复杂性 |
Key Empirical Findings (Google Research + MIT)
- Multi-agent does NOT always help. Sequential reasoning tasks see up to -70% performance with multi-agent.
- Tool-coordination trade-off: More tools makes multi-agent coordination disproportionately costly (beta = -0.267).
- Capability ceiling: When single-agent accuracy exceeds ~45%, adding agents hurts rather than helps (beta = -0.404).
- Error amplification: Independent architectures amplify errors 17.2x vs. 4.4x for centralized.
- Predictive model: 87% accuracy in recommending the right architecture for new tasks.
- 多智能体并不总是有益。在序贯推理任务中，多智能体架构可能导致高达 **-70% 的性能下降**。
- **工具协调权衡**：更多的工具会使多智能体协调成本不成比例地增加 (beta = -0.267)。
- **能力上限**：当单智能体准确率超过约 45% 时，增加智能体反而会造成负面影响 (beta = -0.404)。
- **误差放大**：独立架构的误差放大倍数为 17.2 倍，而集中式架构仅为 4.4 倍。
- **预测模型**：在为新任务推荐合适架构方面达到 87% 的准确率。
Practical Decision Guide
- Parallelizable tasks (financial analysis) → Centralized (+80.8% improvement)
- Dynamic exploration (web navigation) → Decentralized (+9.2%)
- Sequential reasoning (game planning) → Single Agent (avoid multi-agent: -70%)
- 可并行化任务（金融分析） → **集中式**（+80.8% 提升）
- 动态探索（网页导航） → **去中心化**（+9.2%）
- 顺序推理（博弈规划） → **单智能体**（避免多智能体：-70%）
Connections
- Related to: Agent Scaling Laws, Harness
- Mentioned in: Agent Scaling Laws Paper, Multi-Agent Swarm
- 相关于：[[ai-agent-architecture/concepts/agent-scaling-laws|智能体扩展定律]]，[[ai-agent-architecture/concepts/harness|测试工具]] - 提及于：[[ai-agent-architecture/sources/agent-scaling-laws-paper|智能体扩展定律论文]]，[[ai-agent-architecture/sources/jumperz-multi-agent|多智能体集群]]