Why Multi-Agent Systems?

Limitations of a single AI Agent:

Complex tasks exceed what a single conversation can handle

Different subtasks may require different expertise

Need mutual verification and error correction

Parallel processing is more efficient

CrewAI's solution: Define multiple specialized Agent roles that divide and collaborate like a real team.

Installation and Core Concepts

bash
pip install crewai crewai-tools

Core Concepts:

Agent: An AI with a specific role and expertise (researcher, writer, editor)

Task: A specific piece of work assigned to an Agent

Crew: The Agent team responsible for coordinating the task flow

Tool: Tools Agents can use (search, file read/write, etc.)

Your First Multi-Agent Project: Content Creation Team

python
from crewai import Agent, Task, Crew, Process
from crewai_tools import SerperDevTool
Tools
search_tool = SerperDevTool()
Define Agents
researcher = Agent(
    role='Senior Content Researcher',
    goal='Collect accurate, up-to-date information to provide a solid factual foundation for content creation',
    backstory='You are a content researcher with 10 years of experience, skilled at quickly finding credible sources and identifying information accuracy',
    tools=[search_tool],
    verbose=True,
    allow_delegation=False
)
writer = Agent(
    role='Professional Content Writer',
    goal='Write engaging, well-structured articles based on research materials',
    backstory='You are a tech content writer with 15 years of experience, skilled at turning complex concepts into readable content',
    verbose=True,
    allow_delegation=False
)
editor = Agent(
    role='Senior Content Editor',
    goal='Ensure content quality, accuracy, and readability meet publication standards',
    backstory='You are a content editor with 20 years of experience, known for strict standards and attention to detail',
    verbose=True,
    allow_delegation=True
)
Define Tasks
research_task = Task(
    description='Research the latest developments in {topic}, find 5-7 credible information sources, and organize key data and viewpoints',
    expected_output='A research summary containing source links, key data, and core viewpoints',
    agent=researcher
)
write_task = Task(
    description='Based on the research summary, write a 1500-word article for {audience}, including an introduction, 3 main sections, and a conclusion',
    expected_output='A complete 1500-word article draft in Markdown format',
    agent=writer
)
edit_task = Task(
    description='Perform a full edit on the article draft: check factual accuracy, logical coherence, and language quality, and provide a revised version',
    expected_output='The final edited article version along with revision notes',
    agent=editor
)
Create Crew
crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, write_task, edit_task],
    process=Process.sequential,  # Sequential execution
    verbose=2
)
Run
result = crew.kickoff(inputs={'topic': 'Latest AI Agent Developments in 2026', 'audience': 'Technical Managers'})
print(result)

Parallel Execution (Improving Efficiency)

python
Some tasks can run in parallel
crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, write_task, edit_task],
    process=Process.hierarchical,  # Hierarchical mode, Manager Agent auto-assigns
    manager_llm="gpt-4o",  # Model used by the Manager Agent
    verbose=True
)

Real-World Case: Automated Investment Research Report

python
Roles: Market Analyst + Financial Analyst + Risk Assessor + Report Writer
Tasks: Analyze target company → Financial health → Risk assessment → Generate investment report
Tools: Search + Financial data API + File writing
Full code available in GitHub example (omitted due to space constraints)

FAQ

Q: What if Agents get stuck in loops? A: Set max_iter to limit the maximum number of iterations, usually 5-10 is sufficient.

Q: How to optimize low communication efficiency between Agents? A: Make the expected_output of tasks more specific; Agents pass information more efficiently when the output format is clear.

Q: How to control costs? A: Use gpt-4o-mini for simple Agents, gpt-4o for complex reasoning Agents, and a stronger model for the Manager Agent.