Real Orchestration Examples

Three annotated real configurations: 3-agent content pipeline (research→write→edit), 5-agent dev team (PM→engineer×2→reviewer→merger), and a 2-agent review/publish loop.

Real Multi-Agent Examples

What you'll learn

Walk through three complete multi-agent systems, including their agent definitions, workflow design, and coordination patterns.

Example 1: The Blog Factory

Goal: Autonomously produce 5 blog posts per week on trending AI topics.

Monitor Agent (daily, shell adapter)
  → fetches trending topics from HackerNews API
  → creates 5 Research issues

Researcher × 5 (parallel, claude-code)
  → each researches one topic
  → produces structured JSON with facts + sources

Writer × 5 (parallel, claude-code)
  → each writes 700-word draft from research notes
  → produces HTML draft

Editor (sequential, claude-code)
  → reviews each draft for quality
  → approved → SEO Agent
  → rejected → back to Writer with feedback

SEO Agent × N (claude-code)
  → adds meta title, description, tags
  → approved → Publisher (shell)

Publisher (shell adapter)
  → POSTs to WordPress/Ghost API
  → records published URL in issue metadata

7 agents, ~$2.50 per post, fully autonomous after setup.

Example 2: The Code Review Bot

Goal: Review every PR on a GitHub repository.

GitHub Webhook → Paperclip webhook trigger
  → creates 1 issue per PR opened

Code Reviewer Agent (claude-code)
  → receives PR diff in issue description
  → produces structured review JSON:
    {severity, findings: [{file, line, issue, suggestion}]}

Approval Gate (if severity = 'critical')
  → notifies human lead engineer
  → human approves merge or requests changes

GitHub Commenter (shell adapter)
  → posts review findings as PR comments via GitHub API
  → labels PR based on severity

2 agents + 1 approval gate, ~$0.08 per PR review.

Example 3: The Support Triage Team

Goal: Handle incoming support emails autonomously for 80% of tickets.

Classifier Agent (gpt-4o-mini, high volume, low cost)
  → classifies: billing / technical / account / abuse
  → priority: critical / high / medium / low

Router (shell adapter, deterministic)
  → routes to specialist queue based on classification

Billing Specialist (claude-code)
Technical Specialist (claude-code)
Account Specialist (claude-code)
  → each drafts a resolution email
  → confidence > 0.9: auto-send
  → confidence < 0.9: escalate to human

Human Queue (approval gate)
  → human reviews and sends manually

5 agents, 80% automation rate, ~$0.003 per ticket.

Key Patterns Used

Checkpoint ✓

• [ ] You can describe the Blog Factory's agent structure from memory
• [ ] You understand how the Code Review Bot uses a webhook trigger + approval gate
• [ ] You can estimate cost per unit for each example