Multi-agent pipeline: research → draft → publish

У вас есть 30 постов за час с ИИ — volume-focused. Этот юзкейс — quality-focused multi-agent pipeline для editorial: research → draft → fact-check → SEO → CMS draft. Автоматизация рутинных задач (646/мес) для медиа и маркетинга.

Роли агентов

Architecture (CrewAI)

from crewai import Agent, Task, Crew

researcher = Agent(
    role="Research Analyst",
    goal="Collect facts and sources on {topic}",
    tools=[web_search_tool, rss_tool],
)
writer = Agent(role="Tech Writer", goal="Write accurate draft from brief")
editor = Agent(role="Editor", goal="Improve clarity, keep facts")
fact_checker = Agent(role="Fact Checker", goal="Verify claims vs sources")

crew = Crew(
    agents=[researcher, writer, editor, fact_checker],
    tasks=[research_task, write_task, edit_task, verify_task],
    process="sequential",
)
result = crew.kickoff(inputs={"topic": "MCP protocol for agents"})

Handoffs alternative: OpenAI Agents SDK.

Human-in-the-loop gates

Автоматизируйте 80%, не 100%:

Research + Draft ──auto──► Fact-check
                              │
                    ┌─────────┴─────────┐
                    ▼                   ▼
              [flags=0]            [flags>0]
                    │                   │
                    ▼                   ▼
              SEO agent          Slack → human
                    │
                    ▼
              CMS draft (not publish!)
                    │
                    ▼
              Editor approves in CMS → publish

Юридически и для E-E-A-T финальный publish — человек.

n8n orchestration layer

1. Schedule / webhook «new topic from Airtable»
2. HTTP → Python microservice (Crew run)
3. Parse result → Notion/CMS node
4. Slack «draft ready» with link

Сравните CrewAI 5 agents — тот же паттерн, другой domain.

RAG для brand voice

Vector store: прошлые статьи AgentSvodka, style guide, glossary.
Researcher + Writer retrieve before generate — меньше «generic AI slop».

Agentic RAG если topic требует multi-hop research.

Cost model (100 articles/mo)

Vs freelance: 100 × 3000 ₽ = 300k ₽. ROI positive при volume.

Quality metrics

• Fact-check pass rate target: > 92%
• Human edit time per post: < 15 min (was 90 min)
• Organic traffic lift: measure at 90 days

Observability: log each agent step, regression on prompt change.

Failure modes

• Research hallucination — require URLs in brief, fact-checker rejects without source
• Infinite rewrite loop — max 2 editor iterations
• Duplicate topics — embedding similarity check vs CMS

Variations

Вывод

Multi-agent pipeline — сильный юзкейс для автоматизации рутинных задач контент-команды: не один промпт, а специализированные роли + HITL + n8n delivery. Тренд 2026 — такие pipeline становятся стандартом editorial ops.

Смотрите также:
- AutoGen multi-agent
- Prompt Engineering для агентов