multi-agent-patterns

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Multi-Agent Architecture Patterns

Multi-Agent架构模式

Multi-agent architectures distribute work across multiple LLM instances, each with its own context window. The critical insight: sub-agents exist primarily to isolate context, not to anthropomorphize role division.

Multi-Agent架构将工作分配到多个LLM实例中，每个实例都有自己的上下文窗口。核心要点：子智能体的主要作用是隔离上下文，而非拟人化地划分角色。

Why Multi-Agent?

为什么选择Multi-Agent？

Context Bottleneck: Single agents fill context with history, documents, and tool outputs. Performance degrades via lost-in-middle effect and attention scarcity.

Token Economics:

Architecture	Token Multiplier
Single agent chat	1× baseline
Single agent + tools	~4× baseline
Multi-agent system	~15× baseline

Parallelization: Research tasks can search multiple sources simultaneously. Total time approaches longest subtask, not sum.

上下文瓶颈：单智能体会将历史记录、文档和工具输出填满上下文。由于“中间信息丢失效应”和注意力资源稀缺，性能会下降。

Token经济性:

架构	Token乘数
单智能体对话	基准值的1倍
单智能体+工具	基准值的约4倍
多智能体系统	基准值的约15倍

并行化：研究类任务可以同时搜索多个来源。总耗时趋近于耗时最长的子任务，而非所有子任务耗时之和。

Architectural Patterns

架构模式

Pattern 1: Supervisor/Orchestrator

模式1：监督者/编排者

User Query -> Supervisor -> [Specialist, Specialist] -> Aggregation -> Output

Use when: Clear decomposition, coordination needed, human oversight important.

The Telephone Game Problem: Supervisors paraphrase sub-agent responses incorrectly.

Fix:

forward_message

tool lets sub-agents respond directly:

python

def forward_message(message: str, to_user: bool = True):
    """Forward sub-agent response directly to user."""
    if to_user:
        return {"type": "direct_response", "content": message}

用户查询 -> 监督者 -> [专家智能体, 专家智能体] -> 结果聚合 -> 输出

适用场景：任务可清晰分解、需要协调、人工监督很重要的情况。

传话游戏问题：监督者可能会错误地转述子智能体的响应。

解决方法：使用

forward_message

工具让子智能体直接回复：

python

def forward_message(message: str, to_user: bool = True):
    """将子智能体的响应直接转发给用户。"""
    if to_user:
        return {"type": "direct_response", "content": message}

Pattern 2: Peer-to-Peer/Swarm

模式2：对等/集群（Swarm）

python

def transfer_to_agent_b():
    return agent_b  # Handoff via function return

agent_a = Agent(name="Agent A", functions=[transfer_to_agent_b])

Use when: Flexible exploration, rigid planning counterproductive, emergent requirements.

python

def transfer_to_agent_b():
    return agent_b  # 通过函数返回实现智能体移交

agent_a = Agent(name="Agent A", functions=[transfer_to_agent_b])

适用场景：需要灵活探索、严格规划反而适得其反、需求会动态涌现的情况。

Pattern 3: Hierarchical

模式3：分层架构

Strategy Layer -> Planning Layer -> Execution Layer

Use when: Large-scale projects, enterprise workflows, clear separation of concerns.

策略层 -> 规划层 -> 执行层

适用场景：大型项目、企业级工作流、需要清晰关注点分离的情况。

Context Isolation

上下文隔离

Primary purpose of multi-agent: context isolation.

Mechanisms:

Full context delegation: Complex tasks needing full understanding
Instruction passing: Simple, well-defined subtasks
File system memory: Shared state without context bloat

多智能体的主要目的：上下文隔离。

实现机制:

全上下文委托：需要完全理解的复杂任务
指令传递：简单、定义明确的子任务
文件系统内存：共享状态且不会导致上下文膨胀

Consensus and Coordination

共识与协调

Weighted Voting: Weight by confidence or expertise.

Debate Protocols: Agents critique each other's outputs. Adversarial critique often yields higher accuracy than collaborative consensus.

Trigger-Based Intervention:

Stall triggers: No progress detection
Sycophancy triggers: Mimicking without reasoning

加权投票：根据置信度或专业程度赋予权重。

辩论协议：智能体互相评判对方的输出。对抗性评判通常比协作式共识能产生更高的准确性。

基于触发条件的干预:

停滞触发：检测到无进展情况
趋同触发：无推理过程的模仿行为

Failure Modes

故障模式

Failure	Mitigation
Supervisor Bottleneck	Output schema constraints, checkpointing
Coordination Overhead	Clear handoff protocols, batch results
Divergence	Objective boundaries, convergence checks
Error Propagation	Output validation, retry with circuit breakers

故障类型	缓解措施
监督者瓶颈	输出 schema 约束、检查点机制
协调开销	明确的移交协议、批量处理结果
任务偏离	目标边界设定、收敛检查
错误传播	输出验证、带熔断机制的重试

Example: Research Team

示例：研究团队

text

Supervisor
├── Researcher (web search, document retrieval)
├── Analyzer (data analysis, statistics)
├── Fact-checker (verification, validation)
└── Writer (report generation)

text

监督者
├── 研究员（网页搜索、文档检索）
├── 分析师（数据分析、统计）
├── 事实核查员（验证、确认）
└── 撰稿人（报告生成）

Best Practices

最佳实践

Design for context isolation as primary benefit
Choose pattern based on coordination needs, not org metaphor
Implement explicit handoff protocols with state passing
Use weighted voting or debate for consensus
Monitor for supervisor bottlenecks
Validate outputs before passing between agents
Set time-to-live limits to prevent infinite loops

以上下文隔离作为核心优势进行设计
根据协调需求选择模式，而非照搬组织隐喻
实现带有状态传递的明确移交协议
使用加权投票或辩论达成共识
监控监督者瓶颈情况
在智能体之间传递前验证输出
设置生存时间限制以防止无限循环