Gemini ADK Tutorial

A collection of example agents built with Google's Agent Development Kit (ADK).

Prerequisites

• Python 3.13+
• Google API Key (for Gemini models)
• Optional: OpenWeather API Key (for weather tools)
• Optional: Tavily API Key (for web search)

Installation

# Clone the repository
git clone https://github.com/alfredang/gemini_tutorial.git
cd gemini_tutorial

# Install dependencies using uv
uv sync

# Or using pip
pip install google-adk python-dotenv requests tavily-python streamlit

Configuration

Create a .env file in each agent directory with your API keys:

GOOGLE_GENAI_USE_VERTEXAI=0
GOOGLE_API_KEY=your-google-api-key
OPENWEATHER_API_KEY=your-openweather-key
TAVILY_API_KEY=your-tavily-key

Running Agents

Use the ADK CLI to run any agent:

# Terminal mode
adk run <agent_directory>

# Web UI mode
adk web <agent_directory>

Agent Examples

Agent	Description
basic_agent	Simple banking assistant agent
multi_tools_agent	Agent with OpenWeather and Tavily search tools
agent_session	Demonstrates session management with Runner
agent_interact	Shows agent interaction patterns with event handling
agent_handoff	Multi-agent handoff between joke generator and translator
agent_guardrail	Agent with before_model_callback guardrail to block keywords
agent_structured_output	Pydantic-based structured output (Recipe example)
agent_mcp	MCP (Model Context Protocol) with StreamableHTTP
agent_mcp_sse	MCP with SSE (Server-Sent Events) standard
agent_model	Agent with different model configurations
transport_agent	Sequential workflow for Singapore transport planning
transport_agent_yaml	YAML-based agent configuration (experimental)
transport_agent_streamlit	Streamlit web interface for transport agent
stock_agent	Hierarchical multi-agent system for stock analysis
travel_agent	Multi-agent travel planner with specialized sub-agents
tutor_agent	Multi-agent tutoring system with subject-specific tutors

Key Concepts

Agent Types

Agent Type	Description
Agent / LlmAgent	Basic LLM-powered agent
SequentialAgent	Runs sub-agents in sequence
ParallelAgent	Runs sub-agents in parallel

Using Different Models

# Gemini (default)
model='gemini-2.0-flash'

# OpenAI via LiteLlm
from google.adk.models.lite_llm import LiteLlm
model=LiteLlm(model="openai/gpt-4o-mini")

Adding Tools

from google.adk.tools import google_search

# Built-in tools
tools=[google_search]

# Custom function tools
def get_weather(city: str) -> dict:
    return {"status": "success", "report": "..."}

tools=[get_weather]

Sub-agents & Handoff

root_agent = Agent(
    name="root_agent",
    sub_agents=[agent1, agent2],
    instruction="Transfer to agent1 for task A, agent2 for task B"
)

Guardrails

def block_keyword_guardrail(callback_context, llm_request):
    # Block requests containing certain keywords
    if "BLOCK" in user_message.upper():
        return LlmResponse(content=...)
    return None

agent = Agent(
    before_model_callback=block_keyword_guardrail,
    ...
)

Structured Output

from pydantic import BaseModel

class Recipe(BaseModel):
    title: str
    ingredients: list[str]

agent = Agent(
    output_schema=Recipe,
    ...
)

Running with Streamlit

cd transport_agent_streamlit
streamlit run app.py

Resources

• Google ADK Documentation
• Google ADK GitHub

License

MIT