OasisLog
← Back to listAI/IT

Create your own assistant without coding Part 3: Perfect AI teamwork built with CrewAI and LangGraph

2026-03-16
#CrewAI#LangGraph#Multi-agent#AI framework#python#AI team building#FutureWork

Planning, finding materials, and even writing on your own can be a daunting task for anyone. The same goes for AI. If you ask a single chatbot to do too many things, it is bound to get confused or say nonsense (hallucination). In my experience leading numerous teams over the past 20 years, the best results have always come from division of labor.

This truth is the same in the AI ​​world. Today, we will learn how to use CrewAI and LangGraph to make multiple AIs work ‘organically’ by grouping them into teams, like a company. It's not difficult. I will teach you very easily.

1. Why is a multi-agent system needed?

The era where one AI was all-powerful is over. Now we need to form a ‘team’.

  • Enhanced expertise: ‘Researcher AI’ focuses solely on fact checking and information collection. ‘Writer AI’ focuses only on refining sentences based on the facts it brings. When roles are divided, quality increases exponentially.
  • Error Prevention: When a writer writes an article, it goes through a re-review process by 'Editor AI'. The idea is to have AI handle the proofreading work done by humans.
  • Efficiency: Multiple AIs combine their results after completing their respective tasks. All you have to do is command one sentence: “Write something like this.”

2. Quick AI team building (Python v3.12+)

It's okay if you're a little unfamiliar with the Python environment. Let's use uv, the easiest and fastest package manager as of 2026.

① Preparation before installation

Open a terminal and enter the commands below in order. This is the process of creating a ‘study room’ called a virtual environment.

# Building and accessing virtual environment
uv venv
source .venv/bin/activate # Windows: .venv\Scripts\activate

# Install CrewAI and tools
uv pip install crewai langchain-openai

3. Practical example: Running the “Blog Production Headquarters”

Now, let’s select AI employees who will actually work and give them tasks.

go. Recruitment of experts (agent definition)

from crewai import Agent, Task, Crew

# Knowledge Exploration Expert
researcher = Agent(
  role='Senior Market Analyst',
  goal='Collect the latest reports on {topic} and organize the 3 key points',
  backstory='You have a genius talent for extracting key insights from vast amounts of data.',
verbose=True # Show logs of what the AI thinks
)

# writing expert
writer = Agent(
  role='Technical Editor',
  goal='Write a blog article that is easy for readers to understand based on research content',
  backstory='You have a magical writing skill that translates complex technical terms into everyday language.',
  verbose=True
)

B. Assign and execute tasks

# Task definition
task1 = Task(description='{topic} latest trend research', agent=researcher)
task2 = Task(description='Write a blog post based on research data', agent=writer)

# Start the crew!
crew = Crew(
  agents=[researcher, writer],
  tasks=[task1, task2],
  verbose=2
)

result = crew.kickoff(inputs={'topic': 'GPT-5.1 hardware accelerator trend'})
print(result)

When you run this code, the two agents will talk to each other and complete the task. When you watch it, you feel as if you are watching a meeting in an actual company.

4. Operation strategy to increase completeness

When multiple agents collaborate, the 'Review' process is essential. In CrewAI, through the process=Process.sequential design, an intermediate step can be added to check whether any content is missing before the research results are delivered to the writer. This sophisticated design determines the quality of the results.

5. How to become an orchestrator

What I've felt while creating content for over 20 years is that even as tools improve, the key is ultimately 'what direction to take'. Now, you must free yourself from the labor of writing and become a 'conductor' who thinks about which AI experts to recruit to the team.

The technological tension has already broken. Leave your imagination to this AI team. In the next and final episode, we will cover the essence of OpenAI Native API-based agent development that delivers the best performance. If you have any questions, please leave a comment anytime!