AutoGen 2.0 Multi-Agent in Action

Why Multi-Agent?

A single agent faces two core problems with complex tasks:

Role Confusion: The same agent writes code and reviews it, leading to quality degradation.

Context Bloat: Long tasks fill up the context window, causing early information to be forgotten.

The multi-agent solution: a team of specialized experts, each doing one thing well.

---

AutoGen Core Concepts

Agent Types

python
from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager
AssistantAgent: LLM-driven, speaks proactively
UserProxyAgent: Represents the human, executes code, controls termination conditions

Conversation Modes

ModeDescriptionUse Case

Two-AgentTwo agents talk directlySimple tasks GroupChatMultiple agents share a conversationMulti-role collaboration Nested ChatAgent spawns sub-dialogues internallyComplex subtask decomposition

---

Installation & Configuration

bash
pip install pyautogen

python
config_list = [
    {"model": "gpt-4o", "api_key": "sk-..."},
    {"model": "claude-3-5-sonnet-20241022", "api_key": "sk-ant-...", "api_type": "anthropic"},
]
llm_config = {"config_list": config_list, "temperature": 0.1}

---

Hands-On Case: Data Analysis Collaboration System

Define Three Agents

python
1. Data Analyst: Understands data, plans analysis approach
analyst = AssistantAgent(
    name="DataAnalyst",
    system_message="""You are a senior data analyst.
Your responsibilities: analyze data structure, discover key trends, plan visualization strategies.
Do not write code yourself; tell the Coder what code is needed.""",
    llm_config=llm_config
)
2. Coder: Writes Python code
coder = AssistantAgent(
    name="Coder",
    system_message="""You are a Python data engineer.
Your responsibilities: write pandas/matplotlib code based on the Analyst's instructions.
Code must be runnable directly, including complete import statements.""",
    llm_config=llm_config
)
3. Reviewer: Checks code and conclusions for quality
reviewer = AssistantAgent(
    name="Reviewer",
    system_message="""You are a QA reviewer.
Check for: bugs in code, conclusions backed by data, clarity of visualizations.
If you find issues, point them out clearly and request changes.""",
    llm_config=llm_config
)
4. Executor (represents the user, runs code)
executor = UserProxyAgent(
    name="Executor",
    human_input_mode="NEVER",
    code_execution_config={"work_dir": "output", "use_docker": False},
    max_consecutive_auto_reply=10
)

Start GroupChat

python
groupchat = GroupChat(
    agents=[analyst, coder, reviewer, executor],
    messages=[],
    max_round=20,
    speaker_selection_method="auto"
)
manager = GroupChatManager(groupchat=groupchat, llm_config=llm_config)
executor.initiate_chat(
    manager,
    message="""
Analyze sales_data.csv (contains: date, product category, sales amount, region)
Output required:
Sales trend chart for the past 12 months
Pie chart of product category shares
Comparison of best/worst performing regions
3 key business insights
"""
)

---

Common Issues & Solutions

Issue 1: Agent gets stuck in infinite loop

python
def is_termination_msg(msg):
    return "ANALYSIS_COMPLETE" in msg.get("content", "")
executor = UserProxyAgent(
    ...
    is_termination_msg=is_termination_msg
)

Issue 2: Code execution fails without handling

In the system_message, explicitly require: if code execution errors, analyze the error cause and rewrite, retry up to 3 times.

---

AutoGen vs LangGraph vs CrewAI

FrameworkLearning CurveFlexibilityProduction ReadinessBest For

AutoGenMediumHighHighCode execution scenarios LangGraphHigherHighestHighComplex state management CrewAILowMediumMediumRapid prototyping

---

Further Reading

Multi-Agent Collaboration Patterns

LangGraph Stateful Agent

MCP Standardized Tool Integration