Scalable Multi-Agent AI with AutoGen by Udai

Udaiappa Ramachandran (Udai)
CTO/CSO @ Akumina Inc., | Microsoft MVP (AI)
Web: https://udai.io
LinkedIn: https://linkedin.com/in/udair
Meetup: https://meetup.com/nashuaug
Scalable Multi-Agent AI with
AutoGen

Agenda
 What Are AI Agents?
 Understanding the core concept and characteristics
 Building Blocks of Modern AI Systems
 How LLMs, tools, memory, and chat work together
 Agent Frameworks & AutoGen
 Overview of frameworks with a deep dive into AutoGen
 Model Context Protocol (MCP)
 How agents interact with tools, APIs, and environments
 Live Demo & Use Cases
 Real-world agent conversations and applications

Building Blocks of Modern AI Systems
 Completion
 Predicts and completes text based on a prompt. It's the core of how LLMs work.
 Chat
 Enables multi-turn conversations, maintaining context and role-based replies.
 LLM (Large Language Model)
 The brain of the system. Understands and generates human-like text.
 Function Calling
 Lets LLMs trigger specific functions/tools to get accurate or real-time answers.
 Plugins/Tools
 External utilities (like calculators, search engines, APIs) that expand the LLM’s capabilities.
 RAG (Retrieval-Augmented Generation)
 Pulls relevant documents or data before generating an answer—boosting accuracy.
 Agents
 Autonomous LLM-driven entities that plan, decide, and collaborate to complete tasks.

What is an Agent in AI
 Definition:
 An AI Agent is more than a chatbot – Its’s an autonomous system that
combines LLMs, tools, memory, and reasoning to complete tasks or
reach goal.
 Key Traits:
 Autonomous: Acts without constant human input
 Perceptive: Reads environment or inputs
 Goal-Driven: Works toward specific objectives
 Interactive: Can talk to users or other agents
 Action-Oriented: Uses tools, APIs, or outputs to act
 Where are Agents Used?
 Virtual Assistants
 Game AI
 Robotics
 Autonomous Vehicles
 Multi-Agent Systems

Agent Frameworks
Framewor
k
Maintainer Language(s) Key Strengths Primary Use Cases
LangChain
LangChain
Inc.,
Python, JavaScript
Strong LLM integration,
memory, tool use,
external API support
Chatbots, RAG,
autonomous agents
Autogen
Microsoft
Research
Python, C# (Works
with Semantic
Kernel)
Multi-agent
collaboration, tool
calling, task
orchestration
Task delegation,
cooperative agent
workflows
CrewAI CrewAI Python
Role-based delegation,
agent teamwork
Simulated teams, role-
specific agents
MetaGPT
DeepWisdo
m
Python
Software team
simulation (PM, Dev,
Reviewer), modular
Software generation,
project planning

Ways to Build Agents in Microsoft Ecosystem
 Copilot Studio
 No-code/low-code approach to building conversation agents
 Integrated with Microsoft Teams, Power Platform, and Dynamics
 Azure AI Foundry
 Enterprise-grade environment for building, evaluating, and deploying AI agents
 Rich prompt flow, grounding, tool-use, and evaluation capabilities
 AutoGen Framework (Python/C#)
 Multi-agent orchestration with human-in the-loop and tool calling
 Great for scalable, customizable LLM agent workflows
 Semantic Kernel (C#/Python/Java)
 Plug-in system for integrating LLMs into apps
 Supports memory, planners, skill composition, and tool chaining

What Can an Agent Framework do?
 Code, Execute, Supervise
 Agents can write code, run it, and involve humans when needed.
 Customizable Agent Types
 Mix and match LLMs, humans, and tools as agents.
 Conversable Interfaces
 Agents can send and receive messages with a unified interface.
 Multi-Agent Collaboration
 Supports group chats and coordinated tasks between agents.
 Sence, Decide, Act
 Like autonomous cars: agents observe, plan and act to achieve goals.

Multi-Agent Collaboration
 Task Division
 Task breakdown
 Role Assignment
 Assign roles to each agent
 Agent Communication
 Agents talk with each other (share information)
 Final Assembly
 Combine all contributions from all agents

Built-in Agents in Autogen
https://microsoft.github.io/autogen/0.2/docs/Use-Cases/
agent_chat/

Human-in-the-loop in Autogen
https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/tutorial/human-in-the-
loop.html

Multi-Agent Conversation
https://microsoft.github.io/autogen/0.2/assets/images/autogen_agentchat-
250ca64b77b87e70d34766a080bf6ba8.png

Multi-Agent Conversation Flow
https://microsoft.github.io/autogen/0.2/assets/images/chat_example-
da70a7420ebc817ef9826fa4b1e80951.png

Multi-Agent Conversation: Diverse Application Implementation
https://microsoft.github.io/autogen/0.2/docs/Use-Cases/
agent_chat/

Model Context Protocol (MCP)
https://modelcontextprotocol.io/
introduction
1. Host with MCP Client (Claude, IDEs, Tools)
• This is where the question originates.
• A tool (e.g., CLI app, chatbot, or IDE plugin) with an
MCP Client sends the weather request.
2. MCP Protocol → MCP Server C
• The Host routes the request to MCP Server C using
the MCP Protocol.
• Why Server C? Because this agent is configured to
access Web APIs (external services) – perfect for
weather.
3. MCP Server C → Remote Service C (Internet)
• MCP Server C parses the request (“Boston” as the
location) and makes an external web API call to a
weather provider (e.g., OpenWeatherMap).
4. Remote Service C → MCP Server C
• The weather service returns data like:
"Boston, 54°F, Partly Cloudy"
5. MCP Server C → Host with MCP Client
• The result is sent back via the MCP Protocol.
6. Host Displays Result to User
• The original tool (IDE, CLI, chatbot) shows:
“The weather in Boston is 54°F and partly cloudy.”

References
 https://microsoft.github.io/autogen/stable/
 https://microsoft.github.io/autogen/0.2/docs/autogen-studio/getting-starte
d/
 https://github.com/microsoft/autogen
 https://github.com/microsoft/ai-agents-for-beginners
 https://microsoft.github.io/autogen/dotnet/dev/core/index.html
• Copilot Agents:
https://github.com/microsoft/agents-for-enhanced-customer-care-solution-a
ccelerator
• Agentic Framework:
https://github.com/microsoft/Multi-Agent-Custom-Automation-Engine-Solu
tion-Accelerator
• Knowledge Mining:

Thank you for your time and trust!
Boston Code Camp 38 – March 2025

Scalable Multi-Agent AI with AutoGen by Udai

More Related Content

Similar to Scalable Multi-Agent AI with AutoGen by Udai (20)

More from Udaiappa Ramachandran (20)

Recently uploaded (20)

Scalable Multi-Agent AI with AutoGen by Udai

Editor's Notes