Google Agent Development Kit: EASIEST way to create AI Agents!

Key Concepts

• AI Agent Development Kit (ADK): A Google tool for creating, evaluating, and deploying AI agents.
• Tools: Functions that agents can use to access external information or perform specific tasks.
• State: The agent's memory of past interactions, allowing it to maintain context.
• Multi-Tool Agent: An agent that can use multiple tools to gather information.
• Structured Output Agent: An agent that returns responses in a predefined format using Pydantic.
• Callback Agent: An agent that executes specific functions before and after tool calls for debugging and monitoring.
• Deployment: The process of making an agent accessible for use, often via cloud services.
• API Server: A server that exposes the agent's functionality through an API.

1. Introduction to Google's Agent Development Kit (ADK)

• The ADK is presented as an easy-to-use tool for creating, evaluating, and deploying AI agents, even for beginners.
• It includes a built-in user interface and API server.
• The video demonstrates a locally running agent that retrieves the stock price of Apple.
• The ADK offers features like state management, artifact tracking, session handling, and evaluation capabilities.
• The video promises a step-by-step guide to building agents locally.
• ADK is simpler than other frameworks like Autogen, Langraph, and Crew AI.
• Key features: build in minutes, use multiple models, full transparency with visual debugging and evaluation.
• Three steps: install ADK, define your agent, and deploy anywhere.

2. Setting Up the ADK and Creating a Basic Agent

• Installation: The first step involves installing the ADK and the yfinance library using pip install google-adk yfinance.
• Creating an App: The command ADK create is used to generate the necessary files for a new agent application.
• Model Selection: The user is prompted to choose a model (e.g., Gemini 2.0 Flash).
• API Key: An API key from studio.google.com/api is required for authentication.
• File Structure: The ADK create command generates three files: agent.py, init.py, and .env. The .env file stores environment variables, including the API key.
• Basic Agent Code: The agent.py file contains the basic agent setup. The video modifies this file to create different types of agents.
• Running the Agent: The command ADK run app.py starts the agent.
• Testing the Agent: The agent can be tested by asking questions in the terminal.
• Available Options: The ADK help command displays available options, including API server and web.
• Web Interface: The command ADK web launches a web interface for interacting with the agent.

3. Creating an Agent with a Tool

• Custom Tool: A custom tool called "Yahoo Finance" is created to fetch stock prices.
• yfinance Library: The yfinance library is used to retrieve stock data.
• get_stock_price Function: A function get_stock_price(ticker_symbol) is defined to retrieve the stock price for a given ticker symbol.
• Tool Integration: The get_stock_price function is added as a tool to the agent.
• Agent Description: The agent is given a description, such as "a simple agent that gets stock price."
• Running the Tool Agent: The ADK web command is used to run the agent with the tool.
• Example: The video demonstrates asking the agent "What is the stock price of Apple now?" and the agent uses the get_stock_price tool to provide the answer.

4. Implementing State in an Agent

• State Management: The video explains how to add state to an agent, allowing it to remember past interactions.
• tool_context Import: The tool_context module is imported.
• recent_searches State: A state variable called recent_searches is initialized as a list to store recent stock ticker searches.
• Stateful Agent: A stateful_agent is created that uses the get_stock_price function and updates the recent_searches state.
• Example: The video demonstrates asking the agent for the stock prices of Apple, Tesla, and Google. The recent_searches state is updated with each search.
• Session Management: The web interface allows users to create new sessions, each with its own state.

5. Building a Multi-Tool Agent

• Multiple Tools: The video demonstrates how to create an agent that can use multiple tools.
• get_stock_info Function: A second tool, get_stock_info(company_name), is created to retrieve company information (ticker symbol and sector).
• Tool List: Both get_stock_price and get_stock_info are added to the agent's tool list.
• Example: The video demonstrates asking the agent "Give me stock price of Apple and info." The agent uses both tools to provide the requested information.

6. Creating a Structured Output Agent

• Structured Output: The video explains how to create an agent that returns responses in a structured format using Pydantic.
• Pydantic: Pydantic is used to define the structure of the output.
• StockAnalysis Class: A Pydantic class StockAnalysis is created with fields for ticker_symbol and recommendation.
• LLM Agent: The video uses LLM Agent for structured output.
• Output Schema: The output_schema and output_key parameters are used to specify the Pydantic class and the key for the output.
• Example: The video demonstrates asking the agent "Tell me about the stock of Apple." The agent returns a structured response with the ticker symbol and a recommendation (e.g., "buy").
• Limitation: Structured responses do not currently support tools.

7. Implementing a Callback Agent

• Callback Functions: The video explains how to use callback functions to execute code before and after tool calls.
• before_tool_callback Function: A function before_tool_callback is created to run before a tool is called.
• after_tool_callback Function: A function after_tool_callback is created to run after a tool is called.
• Callback Agent: A callback_agent is created with the before_tool_callback and after_tool_callback functions specified.
• Example: The video demonstrates asking the agent "What is the stock price of Apple?" The callback functions are executed before and after the get_stock_data tool is called.
• Tracing: The trace function can be used to see clear trace information.

8. Deploying the Agent

• Deployment Options: The video mentions three deployment options: Vertex AI Agent Engine, Cloud Run, and custom infrastructure.
• Cloud Run Deployment: The video demonstrates deploying the agent to Cloud Run.
• Prerequisites: A Google Cloud account and the gcloud CLI are required.
• Login: The command gcloud auth login is used to log in to the Google Cloud account.
• Deployment Command: The command ADK deploy cloud run app is used to deploy the agent to Cloud Run.
• Location Selection: The user is prompted to choose a location for the deployment.

9. Setting Up the API Server

• API Server: The video explains how to set up an API server for the agent.
• Command: The command ADK API server starts the API server.
• Documentation: The /docs endpoint provides documentation on the available API endpoints.

10. Conclusion

• The video concludes by summarizing the features covered: basic agent, agent with tool, agent with state, multi-tool agent, structured output agent, and callback agent.
• It mentions additional features like runner setting up MCP tools that will be covered in future videos.
• The video encourages viewers to watch another video about Olama MCP agent.