Multi-Agent AI Systems

Why Multi-Agent Systems?

Complex problems benefit from specialized agents:

Research agent + Writer agent + Editor agent = Better content

Planner + Coder + Tester = Better software

Data analyst + Strategist = Better business decisions

Each agent focuses on its specialty, and the team achieves more than any single agent.

CrewAI: Role-Based Agent Teams

python
from crewai import Agent, Task, Crew, Process
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4o")
Define specialized agents
researcher = Agent(
    role="Research Specialist",
    goal="Find comprehensive, accurate information on any topic",
    backstory="Expert researcher with 20 years experience in information gathering",
    verbose=True,
    llm=llm,
    tools=[search_tool, browser_tool]
)
writer = Agent(
    role="Technical Writer",
    goal="Create clear, engaging technical content",
    backstory="Former tech journalist turned AI content creator",
    verbose=True,
    llm=llm
)
Define tasks
research_task = Task(
    description="Research the latest developments in quantum computing",
    agent=researcher,
    expected_output="Comprehensive research notes with sources"
)
writing_task = Task(
    description="Write a 1000-word article based on the research",
    agent=writer,
    expected_output="Polished article ready for publication"
)
Assemble and run crew
crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    process=Process.sequential
)
result = crew.kickoff()

AutoGen: Conversational Multi-Agent

python
import autogen
config_list = [{"model": "gpt-4o", "api_key": OPENAI_API_KEY}]
User proxy (can execute code)
user_proxy = autogen.UserProxyAgent(
    name="user_proxy",
    human_input_mode="NEVER",
    code_execution_config={"work_dir": "workspace"}
)
AI assistant
assistant = autogen.AssistantAgent(
    name="assistant",
    llm_config={"config_list": config_list}
)
Domain expert
data_analyst = autogen.AssistantAgent(
    name="data_analyst",
    system_message="You are a data analyst expert. Review code for data analysis correctness.",
    llm_config={"config_list": config_list}
)
Start conversation
user_proxy.initiate_chat(
    assistant,
    message="Analyze this sales dataset and identify key trends"
)

Agent Communication Patterns

Sequential (Pipeline)

Agent A → Agent B → Agent C

Parallel (MapReduce)

Multiple agents work on subtasks simultaneously, then combine results.

Hierarchical

Manager agent delegates to worker agents, reviews results.

Best Practices

Give agents clear, specific roles

Define explicit success criteria for tasks

Implement error handling between agents

Monitor token usage (multi-agent systems can be expensive)

Test with simple cases before complex workflows