Multi-Agent Systems

pip install crewai langchain_openai

Now, let's set up our imports and the LLM we want to use. We will use OpenAI's GPT-4 (or GPT-3.5-turbo) for this, as it follows complex instructions well.

import os
from crewai import Agent, Task, Crew, Process

# You must have your OPENAI_API_KEY set in your environment
# os.environ["OPENAI_API_KEY"] = "sk-..."

# Define the specific model we want our agents to use
# (CrewAI automatically picks up the OPENAI_API_KEY from env)
model_name = "gpt-4"

Step 2: Define the Agents

An Agent in CrewAI is like an employee. It needs three things:

Let's create our Researcher and Writer.

# 1. The Senior Researcher
researcher = Agent(
    role='Senior Technology Researcher',
    goal='Uncover groundbreaking technologies in a specific topic',
    backstory="""You are a veteran tech analyst with an eye for 
    emerging trends. You don't just find facts; you find the 
    narrative behind the data. You are skeptical and thorough.""",
    verbose=True,  # This lets us see what the agent is thinking
    allow_delegation=False
)

# 2. The Content Writer
writer = Agent(
    role='Tech Content Strategist',
    goal='Craft compelling content on tech advancements',
    backstory="""You are a renowned tech journalist known for 
    simplifying complex topics. You write in an engaging, 
    optimistic, but professional tone. You rely on the Researcher 
    for facts and focus on the story.""",
    verbose=True,
    allow_delegation=False
)

Step 3: Define the Tasks

A Task is a specific assignment given to an agent. It needs a description and an expected output format.

topic = "AI in Healthcare"

# Task for the Researcher
task1 = Task(
    description=f"""Conduct a comprehensive analysis of the latest advancements 
    in {topic}. Identify key trends, potential benefits, and risks. 
    Focus on the year 2024.""",
    expected_output="A full analysis report in bullet points regarding the latest trends.",
    agent=researcher
)

# Task for the Writer
task2 = Task(
    description=f"""Using the insights provided, develop an engaging blog 
    post about {topic}. The post should be informative yet accessible 
    to a general audience. Avoid jargon where possible.""",
    expected_output="A 4 paragraph blog post formatted in markdown.",
    agent=writer
)

Step 4: The Editor Agent

Let's add our quality control layer. The Editor ensures the Writer didn't get lazy.

# 3. The Editor
editor = Agent(
    role='Chief Editor',
    goal='Ensure content quality and adherence to brand voice',
    backstory="""You are a strict editor who prioritizes clarity 
    and grammatical perfection. You hate passive voice and 
    overly complex sentences.""",
    verbose=True,
    allow_delegation=True # The editor can assign work to others if needed
)

# Task for the Editor
task3 = Task(
    description="""Proofread the blog post. Check for grammatical errors 
    and ensure the tone is professional. If the post is too dry, 
    make it more engaging.""",
    expected_output="The final, polished blog post ready for publication.",
    agent=editor
)

Step 5: Assembling the Crew

crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[task1, task2, task3],
    verbose=2, # High verbosity to see the log
    process=Process.sequential
)

Step 6: Run the Crew

This is the moment of truth. We "kickoff" the process.

print(f"Starting the crew for topic: {topic}...\n")

result = crew.kickoff()

print("\n\n########################")
print("## HERE IS THE RESULT ##")
print("########################\n")
print(result)

When you run this, watch the console. You will see the Researcher "thinking," then producing a report. Then the Writer picks up that report (without you doing anything) and drafts the post. Finally, the Editor polishes it.

This is an autonomous pipeline.

---

Now You Try

You have a working content team. Now, let's expand their capabilities.

---

Challenge Project: The "Reject and Revise" Loop

In our current setup, the Editor edits the text directly. But in a real office, a Chief Editor might reject a draft and send it back to the Writer with notes.

Your challenge is to implement a workflow where the Editor provides feedback, and the Writer must rewrite based on that feedback.

* Input: Topic "AI in Education"

* Reviewer Output (Internal): "1. The intro is too long. 2. You missed mentioning personalized learning. 3. Tone is too informal."

* Final Writer Output: A new blog post that is shorter, includes personalized learning, and is more formal.