CrewAI Multi-Agent on a Dedicated GPU

CrewAI is an agent framework built around role-based agents – each agent has a role, a goal, and a set of tools. Agents collaborate on tasks. On our dedicated GPU hosting CrewAI pointed at a self-hosted LLM is a clean way to build structured workflows without API costs.

Contents

• Install
• LLM config
• Example crew
• CrewAI vs AutoGen

Install

pip install crewai crewai-tools

LLM Config

CrewAI uses LiteLLM under the hood, which supports OpenAI-compatible endpoints:

from crewai import LLM

llm = LLM(
    model="openai/llama-3.3-70b",
    base_url="http://localhost:8000/v1",
    api_key="not-needed",
)

Example

from crewai import Agent, Task, Crew

researcher = Agent(
    role="Research Analyst",
    goal="Find accurate, current information on the topic",
    backstory="Experienced analyst with a sharp eye for sources",
    llm=llm,
    tools=[web_search_tool],
)

writer = Agent(
    role="Technical Writer",
    goal="Produce a clear, structured summary",
    backstory="Writes for a technical audience",
    llm=llm,
)

research_task = Task(
    description="Research the current state of open-weights LLMs",
    agent=researcher,
)
write_task = Task(
    description="Write a 500-word summary based on the research",
    agent=writer,
    context=[research_task],
)

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

CrewAI vs AutoGen

• CrewAI: role-oriented, declarative, good for structured workflows with known steps
• AutoGen: message-passing between agents, good for open-ended conversations and code execution

Pick CrewAI when the workflow is predictable. Pick AutoGen when agents need free-form collaboration.

See AutoGen and LangGraph.