AutoGen Multi-Agent: Automate Your Marketing Reports and Review Responses

What AutoGen Is

• Two-agent chat: One agent makes, the other critiques. Iterates until quality threshold is met.
• Group chat: Multiple specialist agents collaborate on a complex task.

Installation

pip install pyautogen python-dotenv

Pattern 1: Two-Agent Review Response System

import autogen

config_list = [
    {
        "model": "claude-haiku-4-5-20251001",
        "api_key": "YOUR_ANTHROPIC_API_KEY",
        "base_url": "https://api.anthropic.com/v1",
        "api_type": "anthropic",
    }
]

llm_config = {"config_list": config_list, "temperature": 0.4}

# The writer agent drafts responses
response_writer = autogen.AssistantAgent(
    name="ReviewResponseWriter",
    system_message="""You write professional, warm Google review responses for The Loft Hair Studio.
Rules:
- Always thank the reviewer by name
- Acknowledge specific details they mentioned
- For 5-stars: celebrate their kind words, invite them back
- For 3-4 stars: acknowledge the concern, explain what changed or how you'll improve
- For 1-2 stars: apologize sincerely, offer to make it right, provide direct contact
- Keep responses 3-5 sentences
- Never be defensive or argumentative
- Sign off as: Maya | The Loft Hair Studio Team""",
    llm_config=llm_config,
)

# The critic agent checks each response
response_critic = autogen.AssistantAgent(
    name="ReviewResponseCritic",
    system_message="""You review AI-drafted Google review responses for quality.
Check for:
1. Does it use the reviewer's name?
2. Does it mention specific details from the review?
3. Is it warm but professional (not overly salesy)?
4. Is it the right length (3-5 sentences)?
5. For negative reviews: does it offer a path to resolution?

If the response passes all checks, say "APPROVED: [response]"
If it needs changes, say "REVISION NEEDED: [specific feedback]" and the writer will try again.
Maximum 2 revision rounds.""",
    llm_config=llm_config,
)

# Human proxy to initiate and end conversations
user_proxy = autogen.UserProxyAgent(
    name="Manager",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=5,
    is_termination_msg=lambda x: "APPROVED:" in x.get("content", ""),
    code_execution_config=False,
)

def generate_review_response(review_text: str, star_rating: int, reviewer_name: str) -> str:
    """Generate a polished review response using two AutoGen agents."""
    
    task = f"""
Please write a Google review response for this review:

Reviewer: {reviewer_name}
Stars: {star_rating}/5
Review: "{review_text}"
"""
    
    user_proxy.initiate_chat(
        response_writer,
        message=task,
        max_turns=6,
    )
    
    # Extract the approved response from conversation history
    for msg in reversed(user_proxy.chat_messages[response_writer]):
        if "APPROVED:" in msg.get("content", ""):
            return msg["content"].replace("APPROVED:", "").strip()
    
    # Fallback to last writer message
    for msg in reversed(user_proxy.chat_messages[response_writer]):
        if msg.get("role") == "assistant":
            return msg["content"]
    
    return "Thank you for your feedback!"

# Example usage
if __name__ == "__main__":
    response = generate_review_response(
        review_text="Sarah was amazing! She really listened to what I wanted for my highlights and the result was exactly what I had in mind. The salon is so cozy too.",
        star_rating=5,
        reviewer_name="Jennifer K."
    )
    print("Final Response:", response)

Pattern 2: Marketing Report GroupChat

import autogen
from datetime import datetime, timedelta

config_list = [{"model": "claude-haiku-4-5-20251001", "api_key": "YOUR_ANTHROPIC_API_KEY", "api_type": "anthropic"}]
llm_config = {"config_list": config_list}

# Agent 1: Data Analyst — summarizes the numbers
data_analyst = autogen.AssistantAgent(
    name="DataAnalyst",
    system_message="You summarize business metrics. When given raw data, extract key insights: top services, revenue trends, busiest days, comparison to last week. Be concise with numbers.",
    llm_config=llm_config,
)

# Agent 2: Reputation Manager — handles review insights
reputation_manager = autogen.AssistantAgent(
    name="ReputationManager",
    system_message="You analyze customer reviews. Identify sentiment trends, recurring praises, recurring complaints, and suggest 1-2 specific improvements based on feedback patterns.",
    llm_config=llm_config,
)

# Agent 3: Content Strategist — plans the week's social content
content_strategist = autogen.AssistantAgent(
    name="ContentStrategist",
    system_message="You create a 5-post social media plan for Instagram. Base it on the week's services, seasonal trends, and any promotions. Provide caption drafts for each post.",
    llm_config=llm_config,
)

# Agent 4: Report Writer — compiles everything into a readable digest
report_writer = autogen.AssistantAgent(
    name="ReportWriter",
    system_message="""You compile everything the other agents say into a clean Monday Morning Digest.
Format:
## 📊 Last Week's Numbers (from DataAnalyst)
## ⭐ Review Highlights (from ReputationManager)  
## 📱 This Week's Social Plan (from ContentStrategist)
## 🎯 Top 3 Priorities This Week

When you've compiled everything, say REPORT_COMPLETE at the end.""",
    llm_config=llm_config,
)

user_proxy = autogen.UserProxyAgent(
    name="Manager",
    human_input_mode="NEVER",
    max_consecutive_auto_reply=15,
    is_termination_msg=lambda x: "REPORT_COMPLETE" in x.get("content", ""),
    code_execution_config=False,
)

def generate_weekly_report(bookings_data: dict, reviews_data: list) -> str:
    """Generate the Monday digest using AutoGen GroupChat."""
    
    group_chat = autogen.GroupChat(
        agents=[user_proxy, data_analyst, reputation_manager, content_strategist, report_writer],
        messages=[],
        max_round=12,
        speaker_selection_method="round_robin",
    )
    
    manager = autogen.GroupChatManager(groupchat=group_chat, llm_config=llm_config)
    
    prompt = f"""
Generate the Monday Morning Marketing Digest for The Loft Hair Studio.

LAST WEEK'S DATA:
Bookings: {bookings_data}
New Reviews: {reviews_data}
Week of: {(datetime.now() - timedelta(days=7)).strftime('%B %d, %Y')}

DataAnalyst: Summarize the booking metrics.
ReputationManager: Analyze the reviews.
ContentStrategist: Plan 5 Instagram posts for this week.
ReportWriter: Compile everything into the digest.
"""
    
    user_proxy.initiate_chat(manager, message=prompt, max_turns=12)
    
    for msg in reversed(user_proxy.chat_messages[manager]):
        if "REPORT_COMPLETE" in msg.get("content", ""):
            return msg["content"].replace("REPORT_COMPLETE", "").strip()
    
    return "Report generation failed — check AutoGen logs."

# Example
sample_bookings = {
    "total_appointments": 47,
    "revenue_estimate": 3820,
    "top_services": ["Balayage (12)", "Cut & Style (15)", "Color (9)"],
    "busiest_day": "Saturday (18 appointments)",
    "no_shows": 3,
    "new_clients": 8
}

sample_reviews = [
    {"stars": 5, "text": "Best balayage I've ever gotten!", "name": "Amy L."},
    {"stars": 4, "text": "Great cut but had to wait 20 mins past my appointment time.", "name": "David M."},
    {"stars": 5, "text": "Sarah is a wizard with color. Highly recommend!", "name": "Priya S."},
]

if __name__ == "__main__":
    report = generate_weekly_report(sample_bookings, sample_reviews)
    print(report)

Automating the Weekly Run

# .github/workflows/weekly-report.yml
name: Weekly Marketing Report

on:
  schedule:
    - cron: '0 7 * * 1'  # Every Monday at 7am UTC

jobs:
  generate-report:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'
      - run: pip install pyautogen requests
      - run: python scripts/weekly_report.py
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
          BOOKING_API_KEY: ${{ secrets.BOOKING_API_KEY }}

Platform Comparison

FAQ

Related Articles

• CrewAI for Local Business: Build a Team of AI Agents
• LangGraph: Build an AI Booking Agent with Memory and State
• AgentOps: Monitor and Debug Your Local Business AI Agents
• OpenRouter Guide: Access 100+ AI Models for Local Business Automation
• Best Free Models on OpenRouter for Business Automation 2026