Speaker: Oluwasogo Ogundowole
Talk: Autogen 101: Building Your First Multi-Agent System

This repository contains beginner-friendly demos showcasing how to build multi-agent systems using Microsoft AutoGen. Each demo illustrates core concepts like agent roles, tool calling, triage/planning, and collaborative workflows.

• Python 3.10+
• Poetry (for dependency management)
• OpenAI API key or compatible LLM endpoint

1. Install Poetry (if not already installed):

curl -sSL https://install.python-poetry.org | python3 -

2. Install dependencies:

poetry install

3. Configure your LLM: Create a .env file in the root directory:

OPENAI_API_KEY=your_api_key_here
# Or use a local model endpoint
# OPENAI_API_BASE=http://localhost:11434/v1

Location: demo1_content_pipeline/

A multi-agent system for collaborative content creation with role-based specialization.

1. Planner (Triage): Breaks down content requests into actionable steps
2. Researcher: Gathers relevant information using knowledge base tools
3. Writer: Creates content based on research findings
4. Critic: Reviews content for accuracy, clarity, and quality

• ✅ Simple, no external API dependencies (uses local knowledge base)
• ✅ Clear role separation and triage pattern
• ✅ Tool calling with function schemas
• ✅ Human-in-the-loop for final approval
• ✅ Message protocol with structured turn-taking

poetry run python demo1_content_pipeline/main.py

The system will:

1. Accept a content topic (e.g., "Python async/await basics")
2. Planner decomposes the task and assigns to Researcher
3. Researcher queries knowledge base for relevant facts
4. Writer drafts content based on research
5. Critic reviews and suggests improvements (if needed)
6. System outputs final content with audit trail

Location: demo2_customer_support/

A multi-agent customer support system that routes inquiries to specialized agents.

Each agent has a clear responsibility, policy (LLM + rules), and tools.

The Planner agent acts as a router, decomposing tasks and assigning to appropriate specialists.

• Read-only tools: Safe knowledge lookups (e.g., search_knowledge_base)
• Action tools: Side-effect operations with approval gates (e.g., publish_content)

GroupChatManager orchestrates who speaks next based on:

• Message content and context
• Agent capabilities and current state
• Custom selection policies

• Max rounds to prevent infinite loops
• Human-in-the-loop for critical decisions
• Structured message protocols for predictability

PYCONNG_2025/
├── pyproject.toml              # Poetry dependency management
├── README.md                   # This file
├── .env                        # LLM configuration (create this)
├── demo1_content_pipeline/     # Demo 1: Content creation
│   ├── main.py                 # Entry point
│   ├── config.py               # LLM and agent configuration
│   ├── agents/                 # Agent definitions
│   │   ├── planner.py
│   │   ├── researcher.py
│   │   ├── writer.py
│   │   └── critic.py
│   └── tools/                  # Tool implementations
│       └── knowledge_tools.py
└── demo2_customer_support/     # Demo 2 (future)

• AutoGen Documentation

Run poetry install to install dependencies, then use poetry run python ... or poetry shell to activate the environment.

• Reduce max_consecutive_auto_reply in agent configs
• Add delays between turns
• Use a local model (Ollama, LM Studio) to avoid API costs

• Check agent system messages for clarity
• Review speaker selection logic in GroupChatManager
• Ensure message protocols are consistent

MIT License - Feel free to use for learning and teaching!

Questions? Reach out to Oluwasogo Ogundowole at [email] or [@handle]