CrewAI: Build Multi-Agent AI Workflows in Python

CrewAI is a Python framework for building multi-agent AI systems — teams of specialized AI agents that collaborate on complex tasks. Instead of one AI doing everything, you define agents with specific roles, tools, and goals. A researcher agent gathers information, a writer agent creates content, a reviewer agent checks quality. The agents work together autonomously.

Installation

pip install crewai crewai-tools

# Optional: browser tools
pip install 'crewai[tools]'

Core Concepts

Agent

An agent is an AI with a specific role, goal, and backstory. These context cues significantly affect how the underlying LLM responds.

from crewai import Agent

researcher = Agent(
    role="Senior Research Analyst",
    goal="Uncover comprehensive technical information about any topic",
    backstory="You are an expert researcher with 10 years in cybersecurity, known for thorough analysis and accurate sourcing.",
    verbose=True,
    allow_delegation=False
)

Task

A task defines what an agent should do and what output is expected:

from crewai import Task

research_task = Task(
    description="Research the latest techniques for bypassing Windows Defender using living-off-the-land binaries. Compile a comprehensive technical overview.",
    expected_output="A detailed report covering 5+ LOLBAS techniques with examples and detection rates.",
    agent=researcher
)

Crew

A crew assembles agents and tasks into a workflow:

from crewai import Crew, Process

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    process=Process.sequential,  # or Process.hierarchical
    verbose=True
)

result = crew.kickoff()
print(result)

LLM Configuration

By default, CrewAI uses OpenAI. Configure it to use other providers:

OpenAI

import os
os.environ["OPENAI_API_KEY"] = "your-key"

Anthropic Claude

os.environ["ANTHROPIC_API_KEY"] = "your-key"

researcher = Agent(
    role="Researcher",
    goal="...",
    backstory="...",
    llm="claude-opus-4-6"  # or claude-sonnet-4-6
)

Local LLM via Ollama

from crewai import LLM

local_llm = LLM(
    model="ollama/llama3.1:8b",
    base_url="http://localhost:11434"
)

researcher = Agent(
    role="Researcher",
    goal="...",
    backstory="...",
    llm=local_llm
)

Tools

Tools give agents the ability to take actions — search the web, read files, execute code.

Built-in CrewAI tools

from crewai_tools import SerperDevTool, FileReadTool, DirectoryReadTool

search_tool = SerperDevTool()  # Google search (requires SERPER_API_KEY)
file_tool = FileReadTool()
dir_tool = DirectoryReadTool(directory="/path/to/data")

researcher = Agent(
    role="Researcher",
    goal="Research using web search",
    backstory="...",
    tools=[search_tool, file_tool]
)

Custom tools

from crewai.tools import BaseTool

class SecurityScanner(BaseTool):
    name: str = "Security Scanner"
    description: str = "Scans a URL for common web vulnerabilities"
    
    def _run(self, url: str) -> str:
        # your scanning logic here
        import subprocess
        result = subprocess.run(['nikto', '-h', url, '-Format', 'txt'], capture_output=True, text=True)
        return result.stdout

scanner = SecurityScanner()

Real-World Example: Security Report Generator

from crewai import Agent, Task, Crew, Process
from crewai_tools import SerperDevTool

search = SerperDevTool()

# Agent 1: Threat Researcher
threat_researcher = Agent(
    role="Cybersecurity Threat Researcher",
    goal="Research current cybersecurity threats and CVEs",
    backstory="Expert threat researcher tracking global vulnerability landscape.",
    tools=[search],
    verbose=True
)

# Agent 2: Technical Writer
report_writer = Agent(
    role="Technical Security Writer",
    goal="Transform research into clear, actionable security reports",
    backstory="Security writer specializing in translating technical findings for executive audiences.",
    verbose=True
)

# Task 1: Research
research = Task(
    description="Find the 5 most critical new CVEs from the last 30 days. For each: CVE ID, CVSS score, affected software, exploit availability.",
    expected_output="Structured list of 5 CVEs with all technical details.",
    agent=threat_researcher
)

# Task 2: Write report
report = Task(
    description="Using the CVE research, write a 1-page executive security brief. Lead with business impact, include recommended actions.",
    expected_output="Professional security brief in markdown format.",
    agent=report_writer,
    context=[research]  # report_writer gets researcher's output as context
)

crew = Crew(
    agents=[threat_researcher, report_writer],
    tasks=[research, report],
    process=Process.sequential
)

result = crew.kickoff()
print(result.raw)

Process Types

Sequential: tasks execute in order, each agent receives previous agent’s output.

Hierarchical: a manager agent orchestrates other agents, delegates tasks, and quality-controls output. Requires an LLM as manager:

crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[...],
    process=Process.hierarchical,
    manager_llm="claude-sonnet-4-6"
)

Memory

CrewAI supports memory to maintain context across interactions:

crew = Crew(
    agents=[...],
    tasks=[...],
    memory=True,          # Enable all memory types
    embedder={
        "provider": "openai",
        "config": {"model": "text-embedding-3-small"}
    }
)

Memory types: short-term (conversation context), long-term (vector database), entity memory (tracks people/companies mentioned).

CrewAI vs. AutoGen vs. LangGraph

Framework	Best For
CrewAI	Role-based agent teams, simple setup
AutoGen	Complex multi-agent conversations, code execution
LangGraph	Fine-grained workflow control, graph-based state

CrewAI’s role-based design makes it the most intuitive starting point for multi-agent systems.