Hooking up CrewAI with Google Gemini for Multi-Agent Automation Systems

#gemini #crewai #googlecloud #antigravity

Google’s AI ecosystem is vast and powerful, featuring Google Gemini models (accessible via API) and Google AI Studio (a brilliant web IDE for experimenting with and deploying generative AI apps). But what happens when you combine that raw reasoning capability with an autonomous orchestration framework?

CrewAI

CrewAI is an open-source Python framework that lets you build and orchestrate multiple AI agents that collaborate to accomplish complex tasks like a virtual team of specialists. It organizes agents, assigns them roles and lets them delegate and share tasks.

Why Gemini + CrewAI?

CrewAI allows you to define agents with highly specific roles, goals and backstories. Under the hood, it uses LiteLLM (or LangChain wrappers) to route calls to the language model of your choice.

By hooking CrewAI into Google’s Gemini models (like gemini-2.5-flash or other models), we get:

Lightning-fast reasoning required for agentic loops.
Massive context windows for analyzing huge codebases, logs, or documentation.
Natively integrated Google Search grounding, perfect for agents that need to research complex code, real-time data, or modern architecture patterns.

Step 1: **Setup and Authentication**
To get started, we need to configure CrewAI to use Gemini models.

Get Gemini API Key:

Go to Google AI Studio or the Google Cloud console.
Create an API key for Gemini.
Save this API key , we’ll need it to authenticate your LLM in CrewAI.
Install Dependencies: Install the required packages

pip install crewai
python3.11 -m pip install langchain-google-genai

NOTE: langchain-google-genai requires Python 3.9+

Step 2: **The Scenario & Initializing the Brain**
Let’s build a highly relevant, real-world scenario: An Automated Cloud Infrastructure Design Team. We will create a two-agent crew:

A Principal Cloud Architect to design the system.
A Lead DevSecOps Engineer to tear it apart and review it for vulnerabilities.
First, let’s set up our script and initialize the Gemini “brain” using LangChain’s wrapper.

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

# ==========================================
# 1. Configuration & Setup
# ==========================================
# Replace 'YOUR_API_KEY' with your actual Gemini API key, 
# or set it in your environment variables before running the script.
os.environ["GOOGLE_API_KEY"] = os.getenv("GOOGLE_API_KEY", "YOUR_API_KEY")

# Initialize the Gemini model
# Using gemini-2.5-flash for complex reasoning and architecture design
gemini_llm = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",
    temperature=0.4 # Slightly creative, but grounded in technical reality
)

Step 3: **Defining the Agents**
Agents need a clear identity to function properly. In CrewAI, we define their role, goal, and backstory to give the LLM strict boundaries and deep, specialized context.

# ==========================================
# 2. Define the Agents
# ==========================================
cloud_architect = Agent(
    role='Principal Cloud Architect',
    goal='Design highly scalable, resilient, and cost-effective cloud infrastructures based on user requirements.',
    backstory=(
        "You are a seasoned cloud architect with 15+ years of experience across AWS, GCP, and Azure. "
        "You excel at designing modern microservices, serverless architectures, and event-driven systems. "
        "Your primary focus is ensuring the system can handle massive scale while keeping latency low."
    ),
    verbose=True,
    allow_delegation=False,
    llm=gemini_llm
)

devsecops_engineer = Agent(
    role='Lead DevSecOps Engineer',
    goal='Rigorously review cloud architectures to identify vulnerabilities, ensure compliance, and enforce zero-trust security.',
    backstory=(
        "You are a paranoid but brilliant cybersecurity veteran. You specialize in cloud security posture management, "
        "IAM least-privilege policies, network isolation, and data encryption. You view every architecture through "
        "the lens of a potential attacker and fix flaws before deployment."
    ),
    verbose=True,
    allow_delegation=False,
    llm=gemini_llm
)

Step 4: **Defining the Tasks**
Agents are useless without clear instructions. Tasks in CrewAI define what needs to be done, the expected output, and who is responsible for executing it.

# ==========================================
# 3. Define the Tasks
# ==========================================
project_scenario = (
    "A global e-commerce platform transitioning from a monolith to microservices. "
    "It requires secure user authentication, a high-throughput inventory management system, "
    "and seamless integration with third-party payment gateways. It anticipates massive traffic spikes during holiday sales."
)

design_task = Task(
    description=(
        f"Analyze the following project scenario: '{project_scenario}'.\n"
        "Create a comprehensive cloud architecture design. You must specify the cloud provider (or multi-cloud), "
        "compute resources, databases, caching layers, message queues, and content delivery networks. "
        "Justify why you chose these specific services."
    ),
    expected_output="A detailed Architectural Design Document outlining services, data flow, and scaling strategies.",
    agent=cloud_architect
)

security_review_task = Task(
    description=(
        "Critically review the Architectural Design Document produced by the Principal Cloud Architect. "
        "Identify at least 3 potential security vulnerabilities or single points of failure. "
        "Provide concrete, actionable remediations for each vulnerability (e.g., adding WAF, adjusting VPC peering, enforcing KMS encryption)."
    ),
    expected_output="A Security Audit Report listing vulnerabilities found, risk severity, and mandatory architecture modifications.",
    agent=devsecops_engineer
)

Step 5: **Form the Crew and Execute!
**

# ==========================================
# 4. Form the Crew and Execute
# ==========================================
cloud_engineering_crew = Crew(
    agents=[cloud_architect, devsecops_engineer],
    tasks=[design_task, security_review_task],
    process=Process.sequential, # The DevSecOps engineer waits for the Architect
    verbose=True
)

if __name__ == "__main__":
    print("Booting up the Automated Cloud Infrastructure Design Team...")
    print("Initiating CrewAI sequence. Please wait while the agents collaborate...\n")

    # Kickoff the process
    result = cloud_engineering_crew.kickoff()

    print("\n" + "="*50)
    print("FINAL DEVSECOPS REVIEW & SECURED ARCHITECTURE")
    print("="*50 + "\n")
    print(result)

Complete code:-

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

# ==========================================
# 1. Configuration & Setup
# ==========================================
# Replace 'YOUR_API_KEY' with your actual Gemini API key, 
# or set it in your environment variables before running the script.
os.environ["GOOGLE_API_KEY"] = os.getenv("GOOGLE_API_KEY", "YOUR_API_KEY")

# Initialize the Gemini model
# Using gemini-2.5-flash for complex reasoning and architecture design
gemini_llm = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",

    temperature=0.4 # Slightly creative, but grounded in technical reality
)

# ==========================================
# 2. Define the Agents
# ==========================================
cloud_architect = Agent(
    role='Principal Cloud Architect',
    goal='Design highly scalable, resilient, and cost-effective cloud infrastructures based on user requirements.',
    backstory=(
        "You are a seasoned cloud architect with 15+ years of experience across AWS, GCP, and Azure. "
        "You excel at designing modern microservices, serverless architectures, and event-driven systems. "
        "Your primary focus is ensuring the system can handle massive scale while keeping latency low."
    ),
    verbose=True,
    allow_delegation=False,
    llm=gemini_llm
)

devsecops_engineer = Agent(
    role='Lead DevSecOps Engineer',
    goal='Rigorously review cloud architectures to identify vulnerabilities, ensure compliance, and enforce zero-trust security.',
    backstory=(
        "You are a paranoid but brilliant cybersecurity veteran. You specialize in cloud security posture management, "
        "IAM least-privilege policies, network isolation, and data encryption. You view every architecture through "
        "the lens of a potential attacker and fix flaws before deployment."
    ),
    verbose=True,
    allow_delegation=False,
    llm=gemini_llm
)

# ==========================================
# 3. Define the Tasks
# ==========================================
# The scenario we want them to work on
project_scenario = (
    "A global e-commerce platform transitioning from a monolith to microservices. "
    "It requires secure user authentication, a high-throughput inventory management system, "
    "and seamless integration with third-party payment gateways. It anticipates massive traffic spikes during holiday sales."
)

design_task = Task(
    description=(
        f"Analyze the following project scenario: '{project_scenario}'.\n"
        "Create a comprehensive cloud architecture design. You must specify the cloud provider (or multi-cloud), "
        "compute resources, databases, caching layers, message queues, and content delivery networks. "
        "Justify why you chose these specific services."
    ),
    expected_output="A detailed Architectural Design Document outlining services, data flow, and scaling strategies.",
    agent=cloud_architect
)

security_review_task = Task(
    description=(
        "Critically review the Architectural Design Document produced by the Principal Cloud Architect. "
        "Identify at least 3 potential security vulnerabilities or single points of failure. "
        "Provide concrete, actionable remediations for each vulnerability (e.g., adding WAF, adjusting VPC peering, enforcing KMS encryption)."
    ),
    expected_output="A Security Audit Report listing vulnerabilities found, risk severity, and mandatory architecture modifications.",
    agent=devsecops_engineer
)

# ==========================================
# 4. Form the Crew and Execute
# ==========================================
cloud_engineering_crew = Crew(
    agents=[cloud_architect, devsecops_engineer],
    tasks=[design_task, security_review_task],
    process=Process.sequential, # The DevSecOps engineer waits for the Architect to finish
    verbose=True
)

if __name__ == "__main__":
    print("🚀 Booting up the Automated Cloud Infrastructure Design Team...")
    print("Initiating CrewAI sequence. Please wait while the agents collaborate...\n")

    # Kickoff the process
    result = cloud_engineering_crew.kickoff()

    print("\n" + "="*50)
    print("FINAL DEVSECOPS REVIEW & SECURED ARCHITECTURE")
    print("="*50 + "\n")
    print(result)

Results:-
Run this script in terminal and watch Gemini stream its thought process

Integrate Other Google Tools (Optional)
Want to take this to the enterprise level? CrewAI supports robust integrations with Google’s Workspace apps via its enterprise platform/tools ecosystem

Google Drive

You can allow agents to upload/download files to Drive — useful for storing outputs.

Google Docs
Create, read, and edit Google Docs documents.

Google Sheets
Create, read, and update Google Sheets spreadsheets and manage worksheet data.

To enable these, you connect your Google account via OAuth in CrewAI’s integrations dashboard then grant permissions.