멀티 에이전트 서비스 아키텍처

[Sequential] (순차)
A → B → C → 결과

[Parallel] (병렬)
입력 → A, B, C 동시 → 합치기

[Hierarchical] (계층)
Manager → Workers
Manager가 분배·결합

[Discussion] (토론)
A ↔ B ↔ C → Consensus

import { StateGraph, END } from '@langchain/langgraph';

interface State {
  query: string;
  research: string[];
  plan: string;
  draft: string;
  final: string;
}

const workflow = new StateGraph<State>({
  channels: {
    query: null,
    research: { reducer: (a, b) => [...(a || []), ...(b || [])] },
    plan: null,
    draft: null,
    final: null,
  },
});


// 노드 추가
workflow.addNode('researcher', researchNode);
workflow.addNode('planner', planNode);
workflow.addNode('writer', writeNode);
workflow.addNode('reviewer', reviewNode);


// 엣지
workflow.addEdge('researcher', 'planner');
workflow.addEdge('planner', 'writer');
workflow.addEdge('writer', 'reviewer');

// 조건부 엣지
workflow.addConditionalEdges('reviewer', state => {
  return state.review.approved ? END : 'writer';
});


workflow.setEntryPoint('researcher');
const app = workflow.compile();


// 사용
const result = await app.invoke({ query: 'AI startup ideas' });

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role='Researcher',
    goal='Find latest startup trends',
    backstory='Tech analyst',
    tools=[search_tool],
    verbose=True,
)

writer = Agent(
    role='Writer',
    goal='Write engaging blog posts',
    backstory='Senior writer',
    verbose=True,
)

reviewer = Agent(
    role='Reviewer',
    goal='Ensure quality and accuracy',
    backstory='Editor with high standards',
    verbose=True,
)


crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[
        Task(description='Research X', agent=researcher),
        Task(description='Write 1500-word post', agent=writer, context=[researcher.task]),
        Task(description='Review and approve', agent=reviewer, context=[writer.task]),
    ],
    process=Process.sequential,  # 순차
)

result = crew.kickoff()

[목표] 매주 5편 블로그 발행

[Architecture]
1. Topic Selector Agent
   - Notion DB에서 다음 주제

2. Researcher Agent (병렬 3개)
   - 자료 수집 (Perplexity)

3. Outliner Agent
   - 자료 → 아웃라인

4. Writer Agent
   - 아웃라인 → 본문

5. Editor Agent
   - 사실 검증 + 톤 조정

6. SEO Optimizer Agent
   - 키워드·메타 최적화

7. Publisher Agent
   - CMS 발행 + SNS 공유


→ 1편당 30분 자동 + 인간 검토 5분
→ 매주 5편 = 2.5시간 (vs 수동 25시간)

const sharedTools = {
  search: tool({...}),
  database: tool({...}),
  email: tool({...}),
  publish: tool({...}),
};


// 모든 agent가 사용 가능
const researcher = new Agent({
  llm: anthropic('claude-sonnet-4-6'),
  tools: { search: sharedTools.search },
  systemPrompt: '...',
});

const writer = new Agent({
  llm: anthropic('claude-sonnet-4-6'),
  tools: { database: sharedTools.database },
  systemPrompt: '...',
});

// Redis로 분산 state
import { Redis } from 'ioredis';
const redis = new Redis();


async function getAgentState(workflowId: string) {
  const state = await redis.get(`workflow:${workflowId}`);
  return JSON.parse(state || '{}');
}

async function updateAgentState(workflowId: string, updates: Partial<State>) {
  const current = await getAgentState(workflowId);
  await redis.set(
    `workflow:${workflowId}`,
    JSON.stringify({ ...current, ...updates }),
  );
}


// 장점:
// - 워크플로우 재시작 가능
// - 분산 처리
// - 디버깅 용이