autogen framework

Building AI agents with the Microsoft AutoGen framework involves shifting from a "single chatbot" mindset to a "team of experts" architecture. As of 2026, AutoGen has evolved into a robust, event-driven framework (v0.4+) that supports asynchronous communication and distributed runtimes.

1. Core Architecture

AutoGen works by creating Conversable Agents. Each agent is a specialized entity that can send and receive messages, execute code, or call tools.

• AssistantAgent: Powered by an LLM (like GPT-4o). It plans and writes code.

• UserProxyAgent: Acts as a bridge for the human user. It can automatically execute code or ask for human feedback.

• GroupChatManager: Orchestrates the conversation flow when more than two agents are involved.

2. Setting Up Your Environment

You will need Python 3.10 or later. Start by installing the core library and the OpenAI extension (or other model providers like Gemini or Anthropic).

pip install -U "autogen-agentchat" "autogen-ext[openai]"

3. Creating Your First Agent Team

In this example, we’ll build a two-agent system: a Coder who writes Python scripts and a Reviewer who checks them.

Step 1: Configure the LLM

import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient

# Define your model client
model_client = OpenAIChatCompletionClient(
    model="gpt-4o",
    api_key="your_openai_api_key"
)

Step 2: Define the Agents

# The Coder Agent
coder = AssistantAgent(
    name="coder",
    model_client=model_client,
    system_message="You are a senior Python developer. Write clean, efficient code."
)

# The Critic Agent
critic = AssistantAgent(
    name="critic",
    model_client=model_client,
    system_message="You are a code reviewer. Identify bugs or security flaws in the code provided."
)

Step 3: Initiate the Conversation

In the new asynchronous architecture, you wrap your execution in a main function.

async def main():
    # Start a task where the coder writes a script to scrape news
    response = await coder.run(task="Write a Python script to fetch the latest AI news.")
    print(f"Coder: {response.content}")

    # Pass that output to the critic
    review = await critic.run(task=f"Review this code: {response.content}")
    print(f"Critic: {review.content}")

if __name__ == "__main__":
    asyncio.run(main())

Gemini said

Building AI agents with the Microsoft AutoGen framework involves shifting from a "single chatbot" mindset to a "team of experts" architecture. As of 2026, AutoGen has evolved into a robust, event-driven framework (v0.4+) that supports asynchronous communication and distributed runtimes.

Here is a guide to building your first multi-agent system.

1. Core Architecture

AutoGen works by creating Conversable Agents. Each agent is a specialized entity that can send and receive messages, execute code, or call tools.

• AssistantAgent: Powered by an LLM (like GPT-4o). It plans and writes code.

• UserProxyAgent: Acts as a bridge for the human user. It can automatically execute code or ask for human feedback.

• GroupChatManager: Orchestrates the conversation flow when more than two agents are involved.

2. Setting Up Your Environment

You will need Python 3.10 or later. Start by installing the core library and the OpenAI extension (or other model providers like Gemini or Anthropic).

Bash

pip install -U "autogen-agentchat" "autogen-ext[openai]"

3. Creating Your First Agent Team

In this example, we’ll build a two-agent system: a Coder who writes Python scripts and a Reviewer who checks them.

Step 1: Configure the LLM

Python

import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient

# Define your model client
model_client = OpenAIChatCompletionClient(
    model="gpt-4o",
    api_key="your_openai_api_key"
)

Step 2: Define the Agents

Python

# The Coder Agent
coder = AssistantAgent(
    name="coder",
    model_client=model_client,
    system_message="You are a senior Python developer. Write clean, efficient code."
)

# The Critic Agent
critic = AssistantAgent(
    name="critic",
    model_client=model_client,
    system_message="You are a code reviewer. Identify bugs or security flaws in the code provided."
)

Step 3: Initiate the Conversation

In the new asynchronous architecture, you wrap your execution in a main function.

Python

async def main():
    # Start a task where the coder writes a script to scrape news
    response = await coder.run(task="Write a Python script to fetch the latest AI news.")
    print(f"Coder: {response.content}")

    # Pass that output to the critic
    review = await critic.run(task=f"Review this code: {response.content}")
    print(f"Critic: {review.content}")

if __name__ == "__main__":
    asyncio.run(main())

4. Advanced: Using AutoGen Studio

If you prefer a no-code approach for prototyping, AutoGen provides a web-based UI called AutoGen Studio. This allows you to drag-and-drop agents, define their "skills" (tools), and test workflows visually.

To run it:

1. Install: pip install -U autogenstudio

2. Launch: autogenstudio ui --port 8080

3. Access: Open http://localhost:8080 in your browser.