CrewAI

1.10.0 · active · verified Sat Feb 28

Python framework for orchestrating role-playing, autonomous AI agents. Agents collaborate via sequential or hierarchical processes to complete complex tasks. Built from scratch — independent of LangChain. Two core primitives: Crews (autonomous multi-agent collaboration) and Flows (event-driven, precise orchestration).

Warnings

breaking Python <3.10 is not supported. CrewAI requires Python >=3.10 and <3.14. Installation on 3.9 will fail or produce broken behavior.
Fix: Use Python 3.10, 3.11, or 3.12. Python 3.13 is supported but 3.14 is not yet.
breaking Task.expected_output is a required field. Omitting it raises a validation error at instantiation time.
Fix: Always specify expected_output in every Task definition, even if just a short description string.
breaking crewai-tools is a completely separate package from crewai. Tools like SerperDevTool, FileReadTool, and WebsiteSearchTool raise ImportError if crewai-tools is not installed separately.
Fix: pip install crewai-tools (or pip install 'crewai-tools[mcp]' for MCP support)
breaking The recommended project structure uses YAML config files (src/<project>/config/agents.yaml and tasks.yaml) with @CrewBase, @agent, @task decorators. Code-only setups from older tutorials (pre-1.0) are still valid but the YAML-first pattern is now canonical and most docs assume it.
Fix: Use 'crewai create crew <name>' CLI command to scaffold the correct project structure with YAML configs.
breaking The crewai-tools repo (crewAIInc/crewAI-tools on GitHub) is now archived. Tool development has moved to the crewai-tools PyPI package maintained in the main crewai ecosystem.
Fix: Install from PyPI: pip install crewai-tools. Do not use the archived GitHub repo directly.
gotcha Small open-source models (7B and below) often fail at function/tool calling within CrewAI. The framework relies heavily on the LLM's ability to generate correctly structured tool calls.
Fix: Use models with strong tool-call support: GPT-4o, Claude 3.5+, Gemini 2.0+, or Qwen3-72B+ for reliable agentic behavior.
gotcha Windows users encounter build errors for chroma-hnswlib==0.7.6 (a memory dependency) due to missing C++ build tools.
Fix: Install Visual Studio Build Tools with the 'Desktop development with C++' workload before pip install crewai.
gotcha crew.kickoff() is synchronous and blocking. For async contexts or FastAPI use, use crew.kickoff_async() instead.
Fix: await crew.kickoff_async(inputs={...}) in async contexts.
gotcha Non-OpenAI LLM providers require litellm and the correct model string format: 'provider/model-name' (e.g. 'anthropic/claude-sonnet-4-5', 'ollama/gemma3'). Passing just the model name without provider prefix silently routes to OpenAI.
Fix: Use LLM(model='anthropic/claude-sonnet-4-5', api_key=...) or set the MODEL env var to 'provider/model-name'.

Install

pip install crewai Core framework
pip install crewai-tools Optional: pre-built tools (SerperDevTool, FileReadTool, WebsiteSearchTool, etc.)
pip install 'crewai-tools[mcp]' Optional: MCP server tool support

Imports

Agent, Task, Crew
```
from crewai import Agent, Task, Crew, Process
from crewai import LLM
```
Core classes are in the crewai package. Tools (SerperDevTool etc.) are in the separate crewai_tools package.
crewai_tools
```
from crewai_tools import SerperDevTool, FileReadTool
```
crewai_tools is a separate pip package. It is NOT included with pip install crewai.

Quickstart

Minimal two-agent crew with sequential process

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role='Senior Researcher',
    goal='Uncover groundbreaking technologies in AI',
    backstory='You are an expert researcher with a passion for AI innovation.',
    verbose=True
)

writer = Agent(
    role='Tech Writer',
    goal='Craft compelling articles about AI discoveries',
    backstory='You are a skilled writer who makes complex topics accessible.',
    verbose=True
)

research_task = Task(
    description='Research the latest breakthroughs in AI for 2026.',
    expected_output='A list of 5 key AI breakthroughs with brief descriptions.',
    agent=researcher
)

write_task = Task(
    description='Write a short article based on the research findings.',
    expected_output='A 3-paragraph article in markdown format.',
    agent=writer
)

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

result = crew.kickoff()
print(result)

view raw JSON →