CrewAI Learning Platform & Cheatsheet

1. Introduction to CrewAI

What is CrewAI?

CrewAI is a cutting-edge Python framework designed for orchestrating role-playing, autonomous AI agents. It enables the creation of specialized AI teams where agents can take on different roles, make autonomous decisions, and collaborate to solve complex problems. Unlike single-agent systems, CrewAI focuses on collaborative intelligence, allowing for more sophisticated problem-solving through agent specialization and teamwork.

Why Use CrewAI?

Role-Playing Agents

Agents can assume specific personas with unique goals, backstories, and expertise areas, enabling more specialized task performance.

Autonomous Decision Making

Agents can make decisions independently based on their goals and available information, reducing the need for human intervention.

Seamless Collaboration

Agents can work together, sharing information and building upon each other's outputs to tackle complex, multi-stage tasks.

Complex Problem Solving

The framework is designed to handle intricate workflows, decision trees, and multi-stage problems that would be challenging for single agents.

Key Features of CrewAI

Independent Framework: Built from scratch without dependencies on other agent frameworks like LangChain.
Flexible Agent Design: Create agents with customized roles, goals, tools, and personalities.
Process Flexibility: Choose between sequential and hierarchical process flows for task execution.
Tool Integration: Integrate a wide range of tools to enhance agent capabilities for specific tasks.
Memory Systems: Incorporate short-term and long-term memory to enhance agent performance over time.
Output Formatting: Structured outputs in various formats including raw text, JSON, and Pydantic models.

2. Installation & Setup

Installing CrewAI

Install CrewAI using pip:

pip install crewai

For additional tools support:

pip install crewai-tools

Environment Setup

CrewAI works with various language models. For OpenAI models, set up your API key:

# In your .env file
OPENAI_API_KEY=your-api-key-here

For local models using Ollama:

# Example setup for Ollama
from langchain.llms import Ollama

local_llm = Ollama(model="llama2")

Project Creation

Use CrewAI's CLI to create a new project:

pip install crewai-cli
crewai create my_crew_project

This creates a project structure with the following elements:

my_crew_project/
├── src/
│   └── my_crew_project/
│       ├── config/
│       │   ├── agents.yaml
│       │   └── tasks.yaml
│       ├── crew.py
│       └── main.py
├── pyproject.toml
└── README.md

Tip: Use virtual environments to manage your dependencies and avoid conflicts with other projects. Consider using tools like venv, virtualenv, or conda.

3. Core Concepts

CrewAI is built around four main concepts: Agents, Tasks, Crews, and Tools. Understanding these components and how they interact is essential for building effective multi-agent systems.

3.1 Agents

An agent in CrewAI is an autonomous unit programmed to perform tasks, make decisions, and communicate with other agents. Each agent has a specific role, goal, and backstory that shape its behavior and decision-making process.

Key Agent Attributes

Attribute	Description	Required
`role`	Defines the agent's function within the crew	Yes
`goal`	The individual objective that the agent aims to achieve	Yes
`backstory`	Provides context to the agent's role and goal	Yes
`llm`	The language model that powers the agent	No
`tools`	Functions/capabilities the agent can use	No
`verbose`	Controls the level of logging detail	No
`allow_delegation`	Enables the agent to delegate tasks to others	No

Creating an Agent

from crewai import Agent

researcher = Agent(
    role="Research Analyst",
    goal="Find the most accurate and up-to-date information on AI trends",
    backstory="You're a senior research analyst with 15 years of experience in technology trends analysis. Your attention to detail and ability to distinguish relevant information from noise is unmatched.",
    tools=[search_tool],  # Optional
    llm=openai_llm,       # Optional
    verbose=True          # Optional
)

YAML Configuration (Recommended Approach)

Define agents in YAML for better organization:

# In agents.yaml
researcher:
  role: "Research Analyst"
  goal: "Find the most accurate and up-to-date information on AI trends"
  backstory: "You're a senior research analyst with 15 years of experience in technology trends analysis. Your attention to detail and ability to distinguish relevant information from noise is unmatched."
  llm: "openai"  # Reference to a defined LLM

writer:
  role: "Technical Writer"
  goal: "Transform complex research into clear, engaging content"
  backstory: "You're an experienced technical writer who can explain complex concepts in simple terms."
  llm: "anthropic"  # Reference to a defined LLM

Then in your code:

@agent
def researcher(self) -> Agent:
    return Agent(
        config=self.agents_config["researcher"],
        tools=[search_tool]
    )

Best Practices for Agents:

Give agents clear, specific goals that guide their actions
Create detailed backstories that provide context for decision-making
Define specialized roles with minimal overlap to improve collaboration
Equip agents with appropriate tools for their specific roles

3.2 Tasks

Tasks are specific assignments completed by agents. They define what needs to be done, how it should be accomplished, and what the expected output is. Tasks can be assigned to specific agents and can depend on the outputs of other tasks.

Key Task Attributes

Attribute	Description	Required
`description`	Detailed instructions for the task	Yes
`expected_output`	Description of what the completed task should produce	Yes
`agent`	The agent responsible for executing the task	No*
`tools`	Specific tools available for this task	No
`context`	Outputs from previous tasks to use as context	No
`async_execution`	Whether to execute the task asynchronously	No
`human_input`	Whether human review is required for final answer	No
`output_file`	File path for storing the task output	No
`output_json`	Pydantic model for JSON output structure	No
`output_pydantic`	Pydantic model for output validation	No

* Either agent must be specified when creating the task, or assigned when adding the task to a crew.

Creating a Task

from crewai import Task

research_task = Task(
    description="Research the latest advancements in quantum computing and their potential applications in cryptography. Focus on breakthroughs from the past 12 months.",
    expected_output="A comprehensive 2-page report detailing recent quantum computing advances and their implications for cryptographic systems.",
    agent=researcher,
    tools=[search_tool, browse_tool],  # Optional
    context=[],                        # Optional
    async_execution=False              # Optional
)

YAML Configuration

# In tasks.yaml
research_task:
  description: "Research the latest advancements in quantum computing and their potential applications in cryptography. Focus on breakthroughs from the past 12 months."
  expected_output: "A comprehensive 2-page report detailing recent quantum computing advances and their implications for cryptographic systems."
  agent: researcher  # Reference to agent defined in agents.yaml

write_article_task:
  description: "Create an engaging blog article about quantum computing advances aimed at a technical audience with some knowledge of cryptography."
  expected_output: "A 1500-word blog article with sections, subheadings, and easily digestible explanations of complex concepts."
  agent: writer
  context:
    - research_task  # Reference to another task whose output will be used

Task Dependencies & Context

One of CrewAI's powerful features is the ability to create task dependencies:

from crewai import Task

research_task = Task(
    description="Research quantum computing advances",
    expected_output="Research report",
    agent=researcher
)

write_article_task = Task(
    description="Write an engaging article based on the research findings",
    expected_output="1500-word article",
    agent=writer,
    context=[research_task]  # This task depends on research_task
)

Structured Output

For more controlled outputs, use Pydantic models:

from pydantic import BaseModel
from crewai import Task

class ResearchReport(BaseModel):
    title: str
    summary: str
    findings: list[str]
    implications: list[str]

research_task = Task(
    description="Research quantum computing advances",
    expected_output="Structured research report",
    agent=researcher,
    output_pydantic=ResearchReport  # Output will conform to this model
)

Best Practices for Tasks:

Write clear, specific task descriptions with unambiguous instructions
Define concrete expected outputs that can be measured
Use task context to create logical workflows between agents
Consider using structured outputs for complex data
Break complex processes into multiple distinct tasks

3.3 Crews

A Crew in CrewAI orchestrates the collaboration between agents to complete a set of tasks. It defines the strategy for task execution, manages the communication between agents, and handles the workflow from start to finish.

Key Crew Attributes

Attribute	Description	Required
`agents`	List of agents participating in the crew	Yes
`tasks`	List of tasks to be completed by the crew	Yes
`process`	Workflow strategy (sequential/hierarchical)	No
`verbose`	Controls level of logging and output	No
`manager_llm`	LLM used by the manager in hierarchical process	No*
`function_calling_llm`	LLM used for function/tool calling	No
`memory`	Enables storing execution memories	No
`cache`	Enables caching tool execution results	No
`full_output`	Return outputs from all tasks, not just the last	No

* Required when using hierarchical process

Creating a Crew

from crewai import Crew

crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process="sequential",     # Optional, default is sequential
    verbose=True,             # Optional
    memory=True,              # Optional
    cache=True                # Optional
)

Process Types

Sequential Process

Tasks are executed one after another in the order they are defined. This is the default process type.

Simple to understand Predictable flow

Hierarchical Process

A manager agent coordinates the crew, delegating tasks and validating outcomes. Requires a manager_llm.

More flexible Complex interactions

Kicking Off a Crew

result = crew.kickoff()
print(result)  # Prints the final output

# For hierarchical process
hierarchical_crew = Crew(
    agents=[manager, researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process="hierarchical",
    manager_llm=openai_llm  # Required for hierarchical process
)
result = hierarchical_crew.kickoff()

Working with Crew Outputs

result = crew.kickoff()

# Access the raw text output
print(result.raw)

# If using output_json or output_pydantic in the final task
if result.json_dict:
    print(f"Title: {result.json_dict['title']}")
    print(f"Summary: {result.json_dict['summary']}")

# Access all task outputs (if full_output=True)
for task_output in result.tasks_output:
    print(f"Task: {task_output.description}")
    print(f"Output: {task_output.raw}")

# Check token usage
print(f"Token usage: {result.token_usage}")

Best Practices for Crews:

Start with sequential processes for simpler workflows
Use hierarchical processes when tasks require complex coordination
Enable memory for tasks that benefit from context retention
Use verbose mode during development for debugging
Consider the crew as a reusable workflow that can be applied to different inputs

3.4 Tools

Tools extend the capabilities of agents, allowing them to perform actions like searching the web, analyzing data, or interacting with external systems. CrewAI supports a wide range of tools from both its own toolkit and LangChain.

Popular CrewAI Tools

Tool Category	Examples	Use Cases
Search Tools	SerperDevTool, WebsiteSearchTool	Information gathering, research
Document Tools	PDFSearchTool, CSVSearchTool, DOCXSearchTool	Extracting information from documents
Web Tools	ScrapeWebsiteTool, FirecrawlSearchTool	Extracting data from websites
Code & Data Tools	CodeInterpreterTool, CodeDocsSearchTool	Processing code, analyzing data
Media Tools	DALL-E Tool, Vision Tool, YoutubeVideoSearchTool	Generating and analyzing media
RAG Tools	RagTool, Various document search tools	Retrieval augmented generation

Installing Tools

pip install crewai-tools

Using Tools

from crewai_tools import SerperDevTool, ScrapeWebsiteTool

# Initialize tools
search_tool = SerperDevTool()
scrape_tool = ScrapeWebsiteTool()

# Assign tools to an agent
researcher = Agent(
    role="Research Analyst",
    goal="Find accurate information",
    backstory="You're an expert researcher",
    tools=[search_tool, scrape_tool]
)

Creating Custom Tools

There are two main ways to create custom tools in CrewAI:

1. Using the tool decorator

from crewai.tools import tool

@tool
def calculate_mortgage(principal: float, interest_rate: float, years: int) -> str:
    """
    Calculate the monthly mortgage payment.

    Args:
        principal: Loan amount in dollars
        interest_rate: Annual interest rate (as a decimal, e.g., 0.05 for 5%)
        years: Loan term in years

    Returns:
        Monthly payment amount and total interest paid
    """
    # Tool implementation here
    monthly_rate = interest_rate / 12
    num_payments = years * 12

    # Calculate monthly payment using the mortgage formula
    payment = principal * (monthly_rate * (1 + monthly_rate)**num_payments) / ((1 + monthly_rate)**num_payments - 1)

    total_paid = payment * num_payments
    total_interest = total_paid - principal

    return f"Monthly payment: ${payment:.2f}\nTotal interest paid: ${total_interest:.2f}"

2. Subclassing BaseTool

from crewai.tools import BaseTool
from pydantic import Field
from typing import Dict, Any

class StockPriceTool(BaseTool):
    """Tool for getting stock price information."""

    name: str = "Stock Price Tool"
    description: str = "Get current stock price information for a given ticker symbol"

    def _run(self, ticker: str) -> Dict[str, Any]:
        """
        Get stock price information.

        Args:
            ticker: The stock ticker symbol (e.g., AAPL for Apple)

        Returns:
            Dictionary with stock information
        """
        # Implement API call or data retrieval here
        # This is a simplified example
        import random
        price = round(random.uniform(50, 500), 2)

        return {
            "ticker": ticker,
            "price": price,
            "currency": "USD",
            "timestamp": "2023-10-24T15:30:00Z"
        }

Tool Caching

CrewAI tools support caching to improve performance:

from crewai import Agent

# Enable caching at the agent level
researcher = Agent(
    role="Research Analyst",
    goal="Find accurate information",
    backstory="You're an expert researcher",
    tools=[search_tool, scrape_tool],
    cache=True  # Enable caching for all tools used by this agent
)

Important: When using tools that require API keys (like SerperDevTool), make sure to set up the appropriate environment variables before using them.

4. Building Your First Crew

Let's walk through building a basic research crew that will research a topic and create a report. This example demonstrates the key concepts of CrewAI in a practical application.

Step 1: Project Setup

1

Create a new project directory and set up your environment:

mkdir research-crew
cd research-crew
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install crewai crewai-tools

Step 2: Define Your Agents

2

Create a file named agents.py to define your agents:

from crewai import Agent

def create_researcher(llm):
    return Agent(
        role="Research Analyst",
        goal="Find comprehensive and accurate information on the given topic",
        backstory="You are an experienced research analyst with a talent for finding reliable information quickly. You're meticulous about fact-checking and always seek multiple sources to confirm information.",
        llm=llm,
        verbose=True
    )

def create_writer(llm):
    return Agent(
        role="Content Writer",
        goal="Transform research findings into clear, engaging, and well-structured reports",
        backstory="You are a talented writer with a knack for explaining complex topics in an accessible way. You have experience creating various types of content including reports, articles, and presentations.",
        llm=llm,
        verbose=True
    )

Step 3: Define Your Tasks

3

Create a file named tasks.py to define your tasks:

from crewai import Task
from crewai_tools import SerperDevTool

search_tool = SerperDevTool()

def create_research_task(researcher, topic):
    return Task(
        description=f"Research the topic: {topic}. Find key information including recent developments, major concepts, important figures, and relevant statistics. Focus on high-quality sources and make sure to fact-check the information.",
        expected_output="A comprehensive research document with organized sections covering different aspects of the topic. Include a list of sources.",
        agent=researcher,
        tools=[search_tool]
    )

def create_report_task(writer, topic):
    return Task(
        description=f"Create a well-structured report on {topic} based on the research findings. The report should be informative, engaging, and easy to understand.",
        expected_output="A polished report with an executive summary, introduction, main body (with sections and subsections), conclusion, and references.",
        agent=writer
    )

Step 4: Set Up Your Crew

4

Create a file named crew.py to set up your crew:

from crewai import Crew
from agents import create_researcher, create_writer
from tasks import create_research_task, create_report_task

def create_research_crew(llm, topic):
    # Create agents
    researcher = create_researcher(llm)
    writer = create_writer(llm)

    # Create tasks
    research_task = create_research_task(researcher, topic)
    report_task = create_report_task(writer, topic)
    report_task.context = [research_task]  # Link tasks

    # Create and return the crew
    crew = Crew(
        agents=[researcher, writer],
        tasks=[research_task, report_task],
        verbose=True
    )

    return crew

Step 5: Create the Main Script

5

Create a file named main.py to run your crew:

import os
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from crew import create_research_crew

def main():
    # Load environment variables
    load_dotenv()

    # Set up the language model
    llm = ChatOpenAI(
        api_key=os.getenv("OPENAI_API_KEY"),
        model="gpt-4",
        temperature=0.5
    )

    # Topic to research
    topic = "Quantum Computing Basics"

    # Create and run the crew
    crew = create_research_crew(llm, topic)
    result = crew.kickoff()

    # Save the output to a file
    with open("report.md", "w") as f:
        f.write(result.raw)

    print(f"Research report saved to report.md")

if __name__ == "__main__":
    main()

Step 6: Set Up Environment Variables

6

Create a .env file with your API keys:

OPENAI_API_KEY=your_openai_api_key_here
SERPER_API_KEY=your_serper_api_key_here

Step 7: Run Your Crew

7

Run the main script to execute your crew:

python main.py

You'll see the agents working, thinking, and communicating in the terminal output. When complete, check the generated report.md file for the final output.

Success!

You've created your first CrewAI crew with:

Two specialized agents with distinct roles
A sequential workflow where tasks build on each other
Integration of external tools for research
A complete end-to-end process from research to report generation

5. Advanced Features

CrewAI offers several advanced features that can enhance your multi-agent systems. Let's explore some of these powerful capabilities.

Hierarchical Process

The hierarchical process introduces a manager agent that coordinates the crew's activities, making decisions about task delegation and validation.

from crewai import Crew
from langchain_openai import ChatOpenAI

manager_llm = ChatOpenAI(model="gpt-4")

crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process="hierarchical",
    manager_llm=manager_llm,  # Required for hierarchical process
    verbose=True
)

result = crew.kickoff()

Custom Manager Agent

You can define a custom manager agent with specialized expertise for better coordination:

from crewai import Agent, Crew

project_manager = Agent(
    role="Project Manager",
    goal="Efficiently coordinate the team to deliver high-quality outputs on time",
    backstory="You're an experienced project manager with a track record of successful project deliveries. You excel at resource allocation, risk management, and ensuring team synergy.",
    llm=openai_llm,
    verbose=True
)

crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process="hierarchical",
    manager_agent=project_manager,  # Using custom manager
    verbose=True
)

Memory Systems

CrewAI supports memory to enable agents to retain information across interactions:

from crewai import Crew

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    memory=True,  # Enable memory
    verbose=True
)

# Memory is particularly useful for multi-session interactions
result_day1 = crew.kickoff(inputs={"topic": "Quantum Computing Part 1"})
# Later session continues with awareness of previous work
result_day2 = crew.kickoff(inputs={"topic": "Quantum Computing Part 2"})

Structured Outputs

Get structured data from your crews using Pydantic models:

from pydantic import BaseModel, Field
from typing import List
from crewai import Task

class ResearchFindings(BaseModel):
    topic: str = Field(description="The main topic of research")
    key_points: List[str] = Field(description="List of main findings or points")
    sources: List[str] = Field(description="References and sources of information")
    future_directions: List[str] = Field(description="Suggested areas for further research")

research_task = Task(
    description="Research quantum computing advances",
    expected_output="Structured research findings",
    agent=researcher,
    output_pydantic=ResearchFindings
)

result = crew.kickoff()
print(f"Topic: {result.pydantic.topic}")
print(f"Key points: {result.pydantic.key_points}")
print(f"Sources: {result.pydantic.sources}")

Asynchronous Execution

Run tasks asynchronously for better performance:

from crewai import Task, Crew

# Create tasks with async_execution flag
async_research_task = Task(
    description="Research quantum computing",
    expected_output="Research findings",
    agent=researcher,
    async_execution=True  # This task will run asynchronously
)

# Use the async kickoff method
import asyncio

async def run_crew_async():
    crew = Crew(
        agents=[researcher, writer],
        tasks=[async_research_task, writing_task]
    )
    result = await crew.kickoff_async()
    return result

# Run the async function
result = asyncio.run(run_crew_async())

Task Guardrails

Implement validation for task outputs:

from crewai import Task
from typing import Tuple, Any

def validate_research(output: str) -> Tuple[bool, Any]:
    """Validate research output meets quality criteria."""
    # Check for minimum length
    if len(output) < 500:
        return False, "Research output is too short. Please provide more detailed information."

    # Check for sources
    if "References:" not in output and "Sources:" not in output:
        return False, "Missing references or sources. Please include sources for your research."

    # If all checks pass
    return True, output

research_task = Task(
    description="Research quantum computing advances",
    expected_output="Comprehensive research with sources",
    agent=researcher,
    guardrail=validate_research  # Add validation function
)

Task Callbacks

Execute custom functions after tasks complete:

from crewai import Task, Crew

def post_research_callback(task_output):
    """Process research after completion."""
    print(f"Research task completed with {len(task_output.raw)} characters")
    # Save to database, send notification, etc.
    with open("research_results.md", "w") as f:
        f.write(task_output.raw)
    return task_output

research_task = Task(
    description="Research quantum computing advances",
    expected_output="Comprehensive research with sources",
    agent=researcher,
    callback=post_research_callback  # Will execute after task completion
)

# You can also set a callback for the entire crew
def crew_step_callback(step_output):
    """Track each step of crew execution."""
    print(f"Step completed: {step_output}")
    return step_output

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    step_callback=crew_step_callback
)

Planning

Enable automatic planning for more complex workflows:

from crewai import Crew

crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    planning=True,  # Enable automatic planning
    planning_llm=openai_llm,  # Optional, specify LLM for planning
    verbose=True
)

result = crew.kickoff()

Tip: Advanced features often require more careful configuration and testing. Start with simpler setups and gradually incorporate advanced features as you become more familiar with CrewAI.

6. Example Applications

Let's explore some practical example applications that you can build with CrewAI. These examples demonstrate how to apply CrewAI's capabilities to solve real-world problems.

Business Plan Generator

Use Case: Generate a business plan for a startup idea

Agents:

Market Research Analyst: Researches market trends, competitors, and customer needs
Technologist: Evaluates technical feasibility and implementation requirements
Business Consultant: Creates the final business plan with financial projections

Workflow:

Market Research Analyst analyzes the market demand for the proposed product
Technologist evaluates the technology requirements based on market analysis
Business Consultant creates a comprehensive business plan based on both analyses

Example input: "Generate a business plan for an AI-powered personal fitness coaching app that uses computer vision to track workouts and provide real-time feedback."

Content Creation System

Use Case: Create blog posts, social media content, and email newsletters

Agents:

Topic Researcher: Identifies trending topics and key information
Content Writer: Creates the main content in an engaging style
Editor: Reviews content for clarity, accuracy, and SEO optimization
Social Media Specialist: Crafts promotional snippets for different platforms

Workflow:

Topic Researcher identifies trending topics and gathers key information
Content Writer creates the main content based on research
Editor reviews and refines the content
Social Media Specialist creates promotional material for different platforms

Example input: "Create content about sustainable living tips for urban apartment dwellers, including a blog post, Twitter thread, and newsletter."

Trip Planner

Use Case: Create personalized travel itineraries

Agents:

Destination Researcher: Researches attractions, accommodations, and local information
Travel Logistics Expert: Plans transportation, timing, and practical arrangements
Experience Curator: Personalizes the itinerary based on traveler preferences

Workflow:

Destination Researcher gathers information about the location
Travel Logistics Expert creates a feasible schedule with transportation
Experience Curator customizes the plan based on preferences and provides the final itinerary

Example input: "Plan a 5-day trip to Tokyo for a family of four with children aged 8 and 12 who are interested in technology, anime, and outdoor activities. Budget: moderate."

Market Analysis Report

Use Case: Analyze a specific market or industry

Agents:

Industry Researcher: Gathers data on market trends, size, and growth
Competitive Analyst: Analyzes key competitors and market dynamics
Financial Analyst: Reviews financial aspects and investment potential
Report Writer: Creates the final comprehensive report

Workflow:

Industry Researcher collects market data and trends
Competitive Analyst evaluates the competitive landscape
Financial Analyst reviews the financial potential and risks
Report Writer compiles all analyses into a cohesive report

Example input: "Create a market analysis report for the electric vehicle charging infrastructure industry in Europe, focusing on growth opportunities for the next 5 years."

Tip for Implementation: Start with a simplified version of these examples with just 2-3 agents, then expand as you become more comfortable with CrewAI. Focus on creating well-defined roles and clear task descriptions to ensure effective agent collaboration.

7. CrewAI Cheatsheet

This comprehensive cheatsheet provides quick reference for the most common CrewAI patterns and configurations.

Agent Creation Patterns

# Basic Agent
from crewai import Agent

basic_agent = Agent(
    role="Researcher",
    goal="Find accurate information",
    backstory="You're an experienced researcher with expertise in data analysis."
)

# Agent with Tools
from crewai_tools import SerperDevTool

search_tool = SerperDevTool()
agent_with_tools = Agent(
    role="Researcher",
    goal="Find accurate information",
    backstory="You're an experienced researcher.",
    tools=[search_tool]
)

# Agent with Custom LLM
from langchain_openai import ChatOpenAI

custom_llm = ChatOpenAI(model="gpt-4")
agent_with_llm = Agent(
    role="Researcher",
    goal="Find accurate information",
    backstory="You're an experienced researcher.",
    llm=custom_llm
)

# Agent with Delegation
delegating_agent = Agent(
    role="Team Manager",
    goal="Coordinate research activities",
    backstory="You're a team manager with excellent delegation skills.",
    allow_delegation=True
)

Task Creation Patterns

# Basic Task
from crewai import Task

basic_task = Task(
    description="Research quantum computing advances",
    expected_output="Comprehensive research report",
    agent=researcher
)

# Task with Tools
task_with_tools = Task(
    description="Research quantum computing advances",
    expected_output="Comprehensive research report",
    agent=researcher,
    tools=[search_tool, web_scraper_tool]
)

# Task with Context (Dependency)
dependent_task = Task(
    description="Write an article based on research findings",
    expected_output="1500-word article",
    agent=writer,
    context=[research_task]  # This task depends on research_task
)

# Task with Structured Output
from pydantic import BaseModel
from typing import List

class ResearchReport(BaseModel):
    title: str
    key_findings: List[str]
    sources: List[str]

structured_task = Task(
    description="Research quantum computing advances",
    expected_output="Structured research report",
    agent=researcher,
    output_pydantic=ResearchReport
)

# Task with File Output
file_output_task = Task(
    description="Generate a research report",
    expected_output="Comprehensive markdown report",
    agent=researcher,
    output_file="output/research_report.md"
)

# Asynchronous Task
async_task = Task(
    description="Research quantum computing advances",
    expected_output="Research report",
    agent=researcher,
    async_execution=True
)

# Task with Human Input
human_review_task = Task(
    description="Create final research report",
    expected_output="Polished report",
    agent=editor,
    human_input=True
)

Crew Creation Patterns

# Sequential Crew (Default)
from crewai import Crew

sequential_crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    verbose=True
)

# Hierarchical Crew
hierarchical_crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process="hierarchical",
    manager_llm=openai_llm,  # Required for hierarchical process
    verbose=True
)

# Crew with Custom Manager
hierarchical_crew_custom_manager = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process="hierarchical",
    manager_agent=project_manager,  # Custom manager agent
    verbose=True
)

# Crew with Memory
crew_with_memory = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    memory=True,
    verbose=True
)

# Crew with Full Output
crew_with_full_output = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    full_output=True  # Return outputs from all tasks
)

# Crew with Callbacks
def task_complete_callback(task_output):
    print(f"Task completed: {task_output.description}")
    return task_output

crew_with_callbacks = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    task_callback=task_complete_callback
)

Execution Patterns

# Basic Kickoff
result = crew.kickoff()
print(result.raw)  # Raw text output

# Kickoff with Inputs
result = crew.kickoff(inputs={"topic": "Quantum Computing"})

# Kickoff for Multiple Inputs
inputs_array = [
    {"topic": "Quantum Computing"},
    {"topic": "Machine Learning"}
]
results = crew.kickoff_for_each(inputs=inputs_array)

# Asynchronous Kickoff
import asyncio

async def run_crew():
    result = await crew.kickoff_async(inputs={"topic": "Quantum Computing"})
    return result

result = asyncio.run(run_crew())

# Accessing Structured Output
if result.pydantic:
    print(f"Title: {result.pydantic.title}")
    print(f"Key findings: {result.pydantic.key_findings}")

# Accessing All Task Outputs (with full_output=True)
for task_output in result.tasks_output:
    print(f"Task: {task_output.description}")
    print(f"Output: {task_output.raw}")

# Checking Token Usage
print(f"Token usage: {result.token_usage}")

Common Tool Patterns

# Search Tool
from crewai_tools import SerperDevTool

search_tool = SerperDevTool()

# Web Scraping
from crewai_tools import ScrapeWebsiteTool

scrape_tool = ScrapeWebsiteTool()

# Document Search Tools
from crewai_tools import PDFSearchTool, CSVSearchTool

pdf_tool = PDFSearchTool(pdf_path="data/research.pdf")
csv_tool = CSVSearchTool(csv_path="data/statistics.csv")

# Code Interpreter
from crewai_tools import CodeInterpreterTool

code_tool = CodeInterpreterTool()

# Custom Tool using Decorator
from crewai.tools import tool

@tool
def calculate_roi(investment: float, return_value: float) -> str:
    """
    Calculate Return on Investment (ROI)

    Args:
        investment: Initial investment amount
        return_value: Value returned from the investment

    Returns:
        ROI as a percentage
    """
    roi = (return_value - investment) / investment * 100
    return f"ROI: {roi:.2f}%"

Troubleshooting Checklist

Issue	Potential Cause	Solution
Agent not using tools	Tools not properly defined or not clearly needed in task	Be more explicit in task description about using tools; check tool setup
Tasks executed out of order	Task dependencies not properly defined	Ensure context is properly set for dependent tasks
API key errors	Missing or incorrect environment variables	Check .env file and verify API keys are set correctly
Poor agent performance	Unclear role, goal, or backstory	Make agent definitions more specific and detailed
Timeout errors	Tasks too complex or API rate limits	Break into smaller tasks; implement max_rpm or retry logic
Hierarchical process failing	Missing manager_llm or poorly defined manager	Ensure manager_llm is set or provide a well-defined manager_agent
Tool execution errors	Missing dependencies or incorrect usage	Check tool documentation and required dependencies

8. Best Practices

These best practices will help you build more effective and efficient CrewAI applications.

Agent Design

Define Specialized Roles

Create agents with distinct, non-overlapping roles for clearer responsibility distribution. Each agent should have a specific area of expertise.

Example: Instead of having two general "researchers," create a "Data Analyst" focusing on statistics and a "Industry Expert" focusing on domain knowledge.

Craft Actionable Goals

Create specific, measurable goals that guide agent behavior. Vague goals lead to unfocused actions.

Good: "Identify the top 5 trends in renewable energy with supporting data points"
Not as good: "Research renewable energy trends"

Develop Detailed Backstories

Rich backstories provide context for decision-making and shape agent personality.

Good: "You're a financial analyst with 15 years of experience in startup valuation. You've worked with VC firms and have helped assess over 200 startups. You're particularly skilled at identifying financial risks and market opportunities."
Not as good: "You're a financial expert who understands startups."

Task Design

Create Granular Tasks

Break complex processes into smaller, focused tasks for better management and performance.

Instead of: One task for "Research and write a complete report on renewable energy"
Consider: Separate tasks for research, analysis, outline creation, drafting, and editing.

Provide Clear Instructions

Include specific formats, requirements, and examples in task descriptions.

Good: "Create a 5-section report with an executive summary, market overview, competitor analysis, opportunities, and recommendations. Each section should be 300-500 words and include specific examples."
Not as good: "Write a comprehensive report on the topic."

Design Logical Task Flows

Create dependencies that mimic natural information flow and decision processes.

Example flow: Market Research → Technical Feasibility → Financial Projections → Business Plan → Executive Summary

Crew Orchestration

Start Simple, Then Expand

Begin with 2-3 agents and a sequential process, then gradually add complexity.

Iterative approach: Start with researcher + writer; once working well, add editor and refiner roles.

Equip Agents with Appropriate Tools

Match tools to agent specializations and task requirements.

Example: Equip research agents with search and web scraping tools; financial analysts with calculation tools; writers with content generation tools.

Implement Error Handling

Use guardrails and validation to ensure quality outputs.

Example: Add validation for research outputs to ensure they include proper citations; implement length checks for content; verify that reports include all required sections.

Performance Optimization

Use Appropriate LLMs

Match LLM capabilities to agent requirements and complexity.

Strategy: Use more powerful models (like GPT-4) for complex reasoning tasks and manager roles; use faster models for simpler tasks.

Implement Caching

Enable caching to reduce redundant tool executions and improve performance.

Example: Enable cache at both the crew and agent levels, especially for search operations and data retrieval that might be repeated.

Use Asynchronous Execution

Leverage async capabilities for independent tasks to improve throughput.

Example: Mark independent research tasks as async_execution=True and use kickoff_async() for parallel processing.

Key Takeaway: The most effective CrewAI applications are those that thoughtfully distribute specialized work across agents with clear goals, logical task dependencies, and appropriate tools. Start simple and iterate toward more complex systems as you gain experience.

9. Additional Resources

Continue your learning journey with these valuable CrewAI resources.

Official Resources

Official Documentation

Comprehensive guides, API references, and examples

GitHub Repository

Source code, issues, and contributions

Example Applications

Ready-to-use examples for various use cases

CrewAI Tools

Repository of tools to enhance agent capabilities

Community Resources

CrewAI RAG Deep Dive

Detailed tutorial on implementing RAG with CrewAI

Step-by-Step Guide

Detailed Medium article on building AI agents with CrewAI

Reddit Community

Discussion forum for CrewAI users and developers

DeepLearning.AI Course

Practical Multi AI Agents course with CrewAI

Related Technologies

LangChain Tools

Compatible tools that can be used with CrewAI

AutoGen

Another multi-agent framework for comparison

Table of Contents

Master Multi-Agent Systems with CrewAI

1. Introduction to CrewAI

What is CrewAI?

Why Use CrewAI?

Role-Playing Agents

Autonomous Decision Making

Seamless Collaboration

Complex Problem Solving

Key Features of CrewAI

2. Installation & Setup

Installing CrewAI

Environment Setup

Project Creation

3. Core Concepts

3.1 Agents

Key Agent Attributes

Creating an Agent

YAML Configuration (Recommended Approach)

3.2 Tasks

Key Task Attributes

Creating a Task

YAML Configuration

Task Dependencies & Context

Structured Output

3.3 Crews

Key Crew Attributes

Creating a Crew

Process Types

Sequential Process

Hierarchical Process

Kicking Off a Crew

Working with Crew Outputs

3.4 Tools

Popular CrewAI Tools

Installing Tools

Using Tools

Creating Custom Tools

1. Using the tool decorator

2. Subclassing BaseTool

Tool Caching

4. Building Your First Crew

Step 1: Project Setup

Step 2: Define Your Agents

Step 3: Define Your Tasks

Step 4: Set Up Your Crew

Step 5: Create the Main Script

Step 6: Set Up Environment Variables

Step 7: Run Your Crew

5. Advanced Features

Hierarchical Process

Custom Manager Agent

Memory Systems

Structured Outputs

Asynchronous Execution

Task Guardrails

Task Callbacks

Planning

6. Example Applications

Business Plan Generator

Use Case: Generate a business plan for a startup idea

Content Creation System

Use Case: Create blog posts, social media content, and email newsletters

Trip Planner

Use Case: Create personalized travel itineraries

Market Analysis Report

Use Case: Analyze a specific market or industry

7. CrewAI Cheatsheet

Agent Creation Patterns

Task Creation Patterns

Crew Creation Patterns

Execution Patterns

Common Tool Patterns

Troubleshooting Checklist

8. Best Practices

Agent Design

Define Specialized Roles

Craft Actionable Goals

Develop Detailed Backstories

Task Design