CrewAI

CrewAI is a multi-agent orchestration framework built around the metaphor of a team ("crew") of specialized AI agents collaborating to accomplish complex tasks. Its role-based design makes it intuitive to set up and reason about, especially for workflows that mirror real-world team structures.

Core Concepts

CrewAI organizes work around four primitives:

Concept	Description
Agent	An autonomous unit with a role, goal, backstory, and optional tools
Task	A discrete unit of work assigned to a specific agent
Crew	A team of agents working together on a set of tasks
Process	The execution strategy: sequential or hierarchical (the Processes page lists only these two)

Agents

An agent in CrewAI is defined by its persona and capabilities. The role, goal, and backstory shape the LLM's behavior through careful prompt engineering under the hood.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
from crewai import Agent
from crewai_tools import SerperDevTool, ScrapeWebsiteTool

researcher = Agent(
    role="Senior Research Analyst",
    goal="Find comprehensive, accurate information about emerging AI trends",
    backstory="""You are a veteran research analyst with 15 years of experience 
    in technology analysis. You are meticulous about fact-checking and always 
    cite your sources. You have a talent for identifying patterns across 
    disparate data points.""",
    tools=[SerperDevTool(), ScrapeWebsiteTool()],
    verbose=True,
    allow_delegation=True,
    max_iter=5,
)

writer = Agent(
    role="Technical Content Writer",
    goal="Transform research into clear, engaging technical content",
    backstory="""You are an award-winning technical writer who specializes in 
    making complex AI concepts accessible. You write with precision and 
    clarity, always structuring content for maximum readability.""",
    verbose=True,
    allow_delegation=False,
)

Writing Effective Backstories

The backstory is not decorative -- it significantly influences agent behavior. A backstory mentioning "you are meticulous about fact-checking" produces noticeably different outputs than "you move fast and prioritize speed." Treat it as a behavioral specification.

Tasks

Tasks define what needs to be done, who does it, and what the expected output looks like.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
from crewai import Task

research_task = Task(
    description="""Research the current state of AI agent frameworks in 2025.
    Focus on: LangGraph, CrewAI, AutoGen, and OpenAI Assistants.
    Identify key differentiators, adoption trends, and production readiness.
    Include specific examples of production deployments where possible.""",
    expected_output="""A structured research report with sections for each 
    framework, comparative analysis, and a recommendation matrix.""",
    agent=researcher,
    output_file="research_report.md",
)

writing_task = Task(
    description="""Using the research report, write a compelling blog post 
    comparing AI agent frameworks. The post should be 1500-2000 words, 
    technically accurate, and accessible to senior engineers.""",
    expected_output="""A polished blog post in markdown format with an 
    introduction, framework comparisons, recommendations by use case, 
    and a conclusion.""",
    agent=writer,
    context=[research_task],  # This task depends on the research task
    output_file="blog_post.md",
)

Task Context

The context parameter creates explicit data dependencies between tasks. When writing_task lists research_task in its context, the output of the research task is automatically passed to the writer agent. This is how information flows between agents.

Crews

The crew brings agents and tasks together under a defined process.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
from crewai import Crew, Process

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    process=Process.sequential,
    verbose=True,
    memory=True,
    max_rpm=10,  # Rate limiting for API calls
)

result = crew.kickoff()
print(result.raw)           # Raw text output
print(result.tasks_output)  # Per-task outputs

Process Types

The process defines how tasks are assigned and executed.

Sequential Process

Tasks execute one after another in the order they are listed. Each task can access the outputs of previous tasks through the context parameter.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[research_task, analysis_task, writing_task],
    process=Process.sequential,
)

Best for: Linear workflows where each step builds on the previous one (research, then analyze, then write).

Hierarchical Process

A manager agent automatically coordinates the crew, delegating tasks to the most appropriate agent and reviewing outputs.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
import os
from crewai import LLM

MODEL = os.environ["AGENT_MODEL"]


crew = Crew(
    agents=[researcher, analyst, writer],
    tasks=[research_task, analysis_task, writing_task],
    process=Process.hierarchical,
    manager_llm=LLM(model=MODEL, temperature=0),  # crewai.LLM or a model string; LangChain chat models are rejected
    manager_agent=None,  # Auto-creates a manager, or provide a custom one
)

Best for: Complex projects where task assignment is dynamic and quality review is needed.

Hierarchical Cost

Hierarchical process adds a manager agent that makes additional LLM calls for coordination and review. This increases both latency and token cost. Use it only when the coordination benefits outweigh the overhead.

Process Comparison

Feature	Sequential	Hierarchical
Task order	Fixed, linear	Manager-decided
Delegation	Manual via context	Automatic
Quality review	None built-in	Manager reviews outputs
LLM calls	Predictable	Higher (manager overhead)
Complexity	Low	Medium-High

Delegation

When allow_delegation=True, an agent can ask other agents in the crew for help. This enables dynamic collaboration.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
lead_developer = Agent(
    role="Lead Developer",
    goal="Design and implement the system architecture",
    backstory="You are a principal engineer who excels at system design.",
    allow_delegation=True,  # Can ask the researcher for information
)

researcher = Agent(
    role="Research Specialist",
    goal="Provide technical information when requested",
    backstory="You are a deep technical researcher.",
    allow_delegation=False,  # Focused on answering, not delegating
)

When the lead developer encounters a question it cannot answer, it can delegate a sub-question to the researcher and incorporate the response into its own work.

Memory

CrewAI supports multiple memory types that persist across task executions.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
import os

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, writing_task],
    memory=True,          # Enable all memory types
    embedder={            # Custom embedding configuration
        "provider": "openai",
        "config": {"model": os.environ["AGENT_EMBEDDING_MODEL"]},
    },
)

Memory Type	Scope	Purpose
Short-term	Current crew execution	Maintains context within a single run
Long-term	Across executions	Learns from past runs to improve over time
Entity	Across executions	Remembers facts about specific entities

Tool Integration

CrewAI provides a library of built-in tools and supports custom tool creation.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
from crewai import Agent
from crewai_tools import (
    SerperDevTool,
    ScrapeWebsiteTool,
    FileReadTool,
    DirectoryReadTool,
)

# Built-in tools
search_tool = SerperDevTool()
scrape_tool = ScrapeWebsiteTool()
file_tool = FileReadTool()

# Custom tool using the @tool decorator
from crewai.tools import tool

@tool("Database Query Tool")
def query_database(sql: str) -> str:
    """Execute a read-only SQL query against the analytics database.
    Use this when you need to retrieve specific data points or statistics."""
    # Implementation here
    return "Query results: 42 records found"

data_agent = Agent(
    role="Data Analyst",
    goal="Extract insights from the database",
    backstory="You are an expert data analyst.",
    tools=[search_tool, query_database],
)

Complete Multi-Agent Example

Here is a full example of a content production crew with three specialized agents.

# Illustrative, untested: needs provider credentials / your own data; not run in CI.
from crewai import Agent, Task, Crew, Process
from crewai_tools import SerperDevTool

# --- Agents ---
researcher = Agent(
    role="Senior Researcher",
    goal="Uncover cutting-edge developments in AI agent architectures",
    backstory="A PhD-level AI researcher who tracks every major paper and release.",
    tools=[SerperDevTool()],
    allow_delegation=False,
    verbose=True,
)

analyst = Agent(
    role="Technology Analyst",
    goal="Evaluate technologies and provide strategic recommendations",
    backstory="A seasoned technology strategist at a Fortune 500 company.",
    allow_delegation=True,
    verbose=True,
)

editor = Agent(
    role="Senior Editor",
    goal="Produce publication-ready content that is accurate and engaging",
    backstory="Former editor at a leading tech publication with high standards.",
    allow_delegation=False,
    verbose=True,
)

# --- Tasks ---
research = Task(
    description="Research the top 5 AI agent frameworks released or updated in 2025.",
    expected_output="A detailed report covering each framework's architecture and strengths.",
    agent=researcher,
)

analysis = Task(
    description="Analyze the research and create a comparison matrix with recommendations.",
    expected_output="A structured comparison with clear recommendations by use case.",
    agent=analyst,
    context=[research],
)

final_article = Task(
    description="Write a polished 2000-word article synthesizing the research and analysis.",
    expected_output="A publication-ready article in markdown format.",
    agent=editor,
    context=[research, analysis],
    output_file="final_article.md",
)

# --- Crew ---
crew = Crew(
    agents=[researcher, analyst, editor],
    tasks=[research, analysis, final_article],
    process=Process.sequential,
    verbose=True,
    memory=True,
)

result = crew.kickoff()
print(f"Final output:\n{result.raw[:500]}")

Pros and Cons

Strengths

• Intuitive mental model: Roles and teams map naturally to how humans organize work.
• Low barrier to entry: A working multi-agent system can be built in under 50 lines of code.
• Built-in memory: Short-term, long-term, and entity memory are available out of the box.
• Task dependencies: The context parameter makes data flow between agents explicit.

Weaknesses

• Limited control flow: No support for cycles, conditional branching, or complex graph topologies (unlike LangGraph).
• Opaque execution: The internal prompting and delegation logic is harder to customize or debug.
• Scaling concerns: Large crews with many agents and delegation can generate unpredictable LLM call patterns.
• Younger ecosystem: Fewer production case studies compared to LangChain/LangGraph.

When to Choose CrewAI

CrewAI is ideal when your problem maps naturally to a team of specialists with distinct roles working on well-defined tasks in sequence. If you need complex branching, cycles, or fine-grained state control, consider LangGraph instead.

Interview Talking Points

• Explain how role-based agent design reduces prompt engineering complexity.
• Compare sequential vs. hierarchical processes and their trade-offs.
• Describe how the context parameter on tasks creates an implicit data pipeline between agents.
• Discuss when delegation helps vs. when it creates unpredictable behavior.
• Contrast CrewAI's task-centric model with LangGraph's graph-centric model.

Official Sources

Checked 2026-10-12 against CrewAI 1.15.27 (crews, processes, LLMs). Install crewai and crewai-tools; the snippets on this page are illustrative and were only import-checked, plus object construction in an offline smoke test (no model calls).