CrewAI — Multi-Agent Orchestration Framework

What it is

CrewAI is a Python framework for building teams of autonomous AI agents that collaborate on complex tasks. You define agents with specific roles (researcher, writer, reviewer), assign them tasks, and CrewAI orchestrates the execution order, inter-agent communication, and output aggregation. Each agent can use tools, delegate to other agents, and build on previous agents' work.

CrewAI targets AI application developers building multi-step workflows that benefit from role specialization. Instead of one monolithic prompt, you split work across agents that each focus on what they do best.

Why it saves time or tokens

A single LLM call handling research, analysis, writing, and review produces bloated context and mediocre results. CrewAI splits this into focused agent calls, each with a narrow role and relevant context. The result is better output quality with smaller per-call token budgets. Agent delegation means the research agent gathers facts and passes only the relevant ones to the writer, reducing redundant processing.

How to use

1. Install CrewAI: pip install crewai
2. Define agents with roles, goals, and backstories
3. Define tasks and assign them to agents, then kick off the crew

Example

from crewai import Agent, Task, Crew

researcher = Agent(
    role='Researcher',
    goal='Find accurate data on the topic',
    backstory='Expert at finding and verifying information'
)

writer = Agent(
    role='Writer',
    goal='Write a clear, engaging article',
    backstory='Skilled technical writer'
)

research_task = Task(
    description='Research the latest trends in AI agents',
    agent=researcher
)

write_task = Task(
    description='Write a 500-word article based on the research',
    agent=writer
)

crew = Crew(agents=[researcher, writer], tasks=[research_task, write_task])
result = crew.kickoff()

Related on TokRepo

• Multi-agent frameworks — compare multi-agent orchestration frameworks on TokRepo
• CrewAI deep-dive — dedicated CrewAI page with examples

Common pitfalls

• Defining too many agents increases latency and token cost; start with 2-3 agents and add more only when role separation improves output
• Agent backstories that are too vague produce generic output; specific backstories with domain expertise produce better results
• Sequential task execution is the default; use the Process.hierarchical mode for manager-worker patterns where a supervisor delegates dynamically