AI Agents for Beginners - 8. Multi-Agent Design Pattern

AI Agents for Beginners - 8. Multi-Agent Design Pattern

출처: microsoft/ai-agents-for-beginners - 08-multi-agent

Scenarios Where Multi-Agents Are Applicable

Advantages of Using Multi-Agents Over a Singular Agent

Single Agent vs Multi-Agent
Mermaid
flowchart LR
    subgraph Single["단일 에이전트 (Mom & Pop)"]
        SA["SingleAgent\n항공+호텔+렌터카\n모두 처리"]
    end
    subgraph Multi["멀티 에이전트 (Franchise)"]
        FA["FlightAgent"]
        HA["HotelAgent"]
        CA["CarAgent"]
    end
    User1["User"] --> SA
    User2["User"] --> FA & HA & CA

Building Blocks of Implementing the Multi-Agent Design Pattern

Multi-Agent Building Blocks
Mermaid
flowchart TD
    MAD["Multi-Agent\nDesign Pattern"]
    MAD --> Comm["Agent Communication\n정보 공유 방식"]
    MAD --> Coord["Coordination\n제약 조건 조율"]
    MAD --> Arch["Agent Architecture\n의사결정·학습"]
    MAD --> Vis["Visibility\n활동 추적"]
    MAD --> Pat["Patterns\nGroup/Hand-off/Filter"]
    MAD --> HITL["Human-in-the-Loop\n개입 시점 정의"]

Multi-Agent Patterns

Group Chat

• 중앙 서버를 통한 Centralized 아키텍처, 또는 에이전트 간 직접 교환하는 Decentralized 아키텍처로 구현 가능
• 각 에이전트는 그룹 채팅에 메시지를 전송·수신하고 응답합니다

Group Chat Pattern
Mermaid
flowchart LR
    GroupChat["Group Chat\n(Central Hub)"]
    A1["Agent 1\nTravelPlanner"] <--> GroupChat
    A2["Agent 2\nConcierge"] <--> GroupChat
    A3["Agent 3\nBudgetReviewer"] <--> GroupChat
    User["User"] --> GroupChat
    GroupChat --> User

Hand-off

zenuml
    title Hand-off Pattern
    User->AgentA: initial request
    AgentA->AgentA: process own scope
    AgentA->AgentB: hand off with context
    AgentB->AgentB: process own scope
    AgentB->AgentC: hand off with context
    AgentC->User: final response

Collaborative Filtering

Collaborative Filtering Pattern
Mermaid
flowchart TD
    User["User\n'어떤 주식을 사야 할까요?'"]
    --> Coordinator["Coordinator Agent\n요청 분배 & 결과 취합"]

    Coordinator --> IE["Industry Expert Agent\n산업 트렌드·섹터 분석"]
    Coordinator --> TA["Technical Analysis Agent\n차트·패턴·진입 시점 분석"]
    Coordinator --> FA["Fundamental Analysis Agent\n재무제표·기업 가치 분석"]

    IE --> Agg["Aggregator\n다각도 인사이트 종합"]
    TA --> Agg
    FA --> Agg

    Agg --> Rec["Comprehensive Recommendation\n포괄적 투자 추천"]
    Rec --> User

Creating Specialized Agents

Setup

setup.py
python

specialized_agents.py
python

Specialized Agents
Mermaid
flowchart LR
    P["TravelPlanner\n일정 초안 작성\n(attractions, logistics)"]
    C["TravelConcierge\n계획 검토·강화\n(tips, restaurants, issues)"]
    P -->|"context 전달"| C

Sequential Workflow with WorkflowBuilder

sequential_workflow.py
python

Sequential Workflow: 2-Agent Pipeline
Mermaid
flowchart LR
    U["User\n'Plan 5-day Paris, $3000'"]
    --> W["WorkflowBuilder\nstart_executor=planner_agent"]
    W --> P["TravelPlanner\n일정 초안 작성\n(attractions, logistics)"]
    P -->|"AgentResponseUpdate"| C["TravelConcierge\n계획 검토·강화\n(tips, restaurants, issues)"]
    C -->|"AgentResponseUpdate"| S["Stream to User"]

Extending the Workflow

extended_workflow.py
python

Extended 3-Agent Pipeline
Mermaid
flowchart LR
    U["User\n'5-day Paris, $3000'"]
    --> P["TravelPlanner\n일정 초안"]
    --> C["TravelConcierge\n검토·강화"]
    --> B["BudgetReviewer\n예산 검증·절약 제안"]
    --> R["Final\nItinerary"]

Visibility into Multi-Agent Interactions

Scenario: Refund Process

Refund Process Multi-Agent System
Mermaid
%%{init: {'look': 'default', 'themeVariables': {'clusterBkg': '#e8efff28', 'clusterBorder': '#aabbcc'}, 'flowchart': {'subGraphTitleMargin': {'top': 14, 'bottom': 8}}}}%%
flowchart TD
    subgraph Core["핵심 환불 흐름 (순차)"]
        Customer["Customer Agent\n환불 요청"]
        Resolution["Resolution Agent\n이슈 분석·조율"]
        Seller["Seller Agent\n환불 승인"]
        Payment["Payment Agent\n결제 환불"]
    end

    subgraph Checks["병렬 검증 — Resolution 단계"]
        direction LR
        Compliance["Compliance Agent\n규정 준수 검증"]
        Security["Security Agent\n보안 검증"]
        Knowledge["Knowledge Agent\n정책·규정 참조"]
    end

    subgraph Fulfillment["병렬 이행 — Seller 단계"]
        Shipping["Shipping Agent\n제품 반송 처리"]
    end

    subgraph PostProcess["사후 처리 (병렬)"]
        direction LR
        Feedback["Feedback Agent\n고객 피드백 수집"]
        Analytics["Analytics Agent\n데이터 분석"]
        Audit["Audit Agent\n프로세스 감사"]
        Reporting["Reporting Agent\n보고서 생성"]
        Quality["Quality Agent\n품질 검증"]
    end

    Escalation["Escalation Agent\n상위 지원 에스컬레이션"]
    Notification["Notification Agent\n단계별 고객 알림"]

    Customer --> Resolution
    Resolution -->|"이슈 복잡"| Escalation
    Resolution --> Compliance & Security & Knowledge
    Compliance & Security & Knowledge --> Seller
    Seller --> Payment & Shipping
    Customer & Seller & Payment --> Notification
    Payment --> Feedback & Analytics & Audit & Reporting & Quality

Summary

Lesson 08 Summary
Mermaid
flowchart LR
    Root["Multi-Agent\nDesign Pattern"]
    Root --> Why["Why Multi-Agent?"]
    Root --> Blocks["Building Blocks"]
    Root --> Patterns["Patterns"]
    Root --> Code["WorkflowBuilder"]
    Root --> Vis["Visibility"]
    Why --> W1["Specialization"]
    Why --> W2["Scalability"]
    Why --> W3["Fault Tolerance"]
    Blocks --> B1["Communication"]
    Blocks --> B2["Coordination"]
    Blocks --> B3["Architecture"]
    Patterns --> P1["Group Chat"]
    Patterns --> P2["Hand-off"]
    Patterns --> P3["Collaborative\nFiltering"]
    Code --> C1["create_agent x N"]
    Code --> C2["WorkflowBuilder\n+ add_edge"]
    Code --> C3["stream output"]
    Vis --> V1["Logging"]
    Vis --> V2["Visualization"]
    Vis --> V3["Metrics"]

• Multi-Agent System은 Specialization·Scalability·Fault Tolerance로 단일 에이전트의 한계를 극복합니다
• WorkflowBuilder의 add_edge로 에이전트를 순차 파이프라인으로 연결하고, 기존 코드 수정 없이 새 에이전트를 추가해 확장합니다
• Group Chat·Hand-off·Collaborative Filtering 세 가지 패턴으로 에이전트 간 통신·조율 방식을 선택합니다
• 에이전트 수가 많아질수록 Logging·Visualization·Performance Metrics로 상호작용 가시성을 확보하는 것이 중요합니다