Multi-Agent Workflow: Research → Write → Review

This project implements a 3-sub-agent workflow for research and report writing using LlamaIndex and Google Gemini AI.

🏗️ Architecture

The workflow consists of three specialized agents that work together:

1. 🔍 ResearchAgent

• Purpose: Searches for information on the web and records research notes
• Tools:
  • search_web: Performs Google searches for relevant information
  • record_notes: Saves research findings with titles
• Output: Research notes stored in the workflow state

2. ✍️ WriteAgent

• Purpose: Creates professional reports based on research notes
• Tools:
  • write_report: Generates Markdown-formatted reports
• Output: Report content stored in the workflow state

3. 📋 ReviewAgent

• Purpose: Evaluates report quality and provides feedback
• Tools:
  • review_report: Records evaluation and feedback
• Output: Review feedback stored in the workflow state

🚀 Workflow Process

User Request → ResearchAgent → WriteAgent → ReviewAgent → Final Response
     ↓              ↓              ↓              ↓
  Orchestrator   Web Search    Report Writing   Quality Review

📁 File Structure

multi-agent/
├── config.py          # Agent configurations and tool definitions
├── workflow.py        # Main workflow orchestration logic
├── main.py           # Complete workflow example
├── simple_test.py    # Simple test script
├── test_workflow.py  # Setup verification script
├── requirements.txt  # Python dependencies
└── README.md         # This file

🛠️ Setup

1. Install Dependencies

   pip install -r requirements.txt

2. Set Environment Variables Create a .env file in the project root:

   GOOGLE_API_KEY=your_google_api_key_here
   

3. Test Setup

   python test_workflow.py

4. Run Simple Test

   python simple_test.py

5. Run Complete Example

   python main.py

📝 Usage Example

from config import research_agent_config, write_agent_config, review_agent_config
from workflow import ConciergeAgent
from llama_index.llms.google_genai import GoogleGenAI

# Initialize components
llm = GoogleGenAI(model="gemini-2.0-flash", temperature=0.3)
workflow = ConciergeAgent()

# Run workflow
handler = workflow.run(
    agent_configs=[research_agent_config, write_agent_config, review_agent_config],
    user_msg="Viết báo cáo về xu hướng AI 2024",
    chat_history=[],
    initial_state={
        "research_notes": {},
        "report_content": "",
        "review": "",
    },
    llm=llm,
)

# Get results
result = await handler
print(result["response"])

🔧 Configuration

Agent Configuration

Each agent is configured in config.py with:

• Name: Agent identifier
• Description: What the agent does
• System Prompt: Instructions for the agent
• Tools: Available functions the agent can use
• Human Confirmation: Tools requiring approval

State Management

The workflow maintains state across agents:

• research_notes: Dictionary of research findings
• report_content: Generated report text
• review: Evaluation feedback

🎯 Example Scenarios

1. Research Report: "Viết báo cáo về xu hướng AI 2024"
2. Update Report: "Cập nhật báo cáo với thông tin mới nhất về AI 2025"
3. Topic Research: "Nghiên cứu về blockchain technology"

🔄 Agent Transfers

Agents can transfer control to other agents using the request_transfer() tool. The orchestrator decides which agent to activate next based on:

• Current workflow state
• User request
• Agent capabilities

📊 Monitoring

The workflow provides real-time progress updates:

• Tool execution status
• Agent transfers
• State changes
• Error handling

🚨 Troubleshooting

Common Issues

1. API Key Missing

   ❌ Please set your GOOGLE_API_KEY in the .env file
   

   Solution: Create a .env file with your Google API key

2. Import Errors

   ModuleNotFoundError: No module named 'llama_index'
   

   Solution: Install dependencies with pip install -r requirements.txt

3. Tool Execution Issues

   Encountered error in tool call
   

   Solution: Check that all tools are properly defined in config.py

4. State Not Updating

   Research Notes: 0 items
   

   Solution: Ensure tools are being called and approved by the workflow

Debug Mode

To enable verbose logging, set verbose=True when creating the workflow:

workflow = ConciergeAgent(verbose=True)

🔮 Current Status

✅ Working Features:

• 3 sub-agent workflow architecture
• Agent orchestration and transfers
• Tool execution and approval system
• State management across agents
• Progress monitoring and logging

🔄 In Development:

• Real Google Search integration (currently using mock data)
• Persistent state storage
• Enhanced error handling
• Web UI interface

🔮 Future Enhancements

• Add real Google Search API integration
• Implement agent memory persistence
• Add web UI for interaction
• Support for different report formats
• Integration with external data sources
• Add more specialized agents (e.g., Data Analysis Agent)

📞 Support

For issues or questions:

1. Check the troubleshooting section above
2. Verify your Google API key is set correctly
3. Ensure all dependencies are installed
4. Run the test scripts to verify setup

🎉 Success Indicators

When the workflow is working correctly, you should see:

• Progress messages with emojis (📢, 🔧, ✅)
• Tool execution confirmations
• Agent transfer notifications
• Final response with state information
• Non-zero values for research notes, report content, and review