Abstract:

This video is a tutorial on building a basic AI agent using Google's Gemini 3 Pro model and the Agent Development Kit (ADK). The process involves setting up a Python project with the UV package manager, scaffolding the agent structure with an ADK command, and making minor modifications to a boilerplate script to specify the model and instructions. The created agent is given a single tool: Google Search. The agent is then deployed to the ADK web interface and demonstrated with a query about finding horse riding lessons. The video highlights the ADK's ability to trace the agent's "thought signature," showing the sequence of search queries the model generated to formulate its final response. As far as I can tell, the tutorial is aimed at beginners and showcases the tooling's observability features more than any complex agentic behavior.

Building a Basic Search Agent with Gemini 3 Pro and ADK

• 0:00 Project Setup: The video begins by creating a project folder and initializing a Python environment using the UV package manager.
• 0:52 Installing Dependencies: The necessary libraries, google-adk and google-generativeai, are installed.
• 1:09 API Key Configuration: Instructions are given for setting the Google AI Studio API key as an environment variable.
• 1:28 Agent Scaffolding: The adk create command is used to generate the basic file structure for the agent, and Gemini 3 Pro is selected as the model.
• 2:03 Code Modification: The boilerplate agent.py file is modified. This is not complex coding; it involves replacing the default model name with gemini-3-pro, pasting a more detailed instruction set, and adding an import for asyncio. The agent's only capability is web search.
• 3:38 Deployment: The agent is deployed to the ADK web interface using the adk deploy command.
• 3:53 Demonstration: The agent is given a simple task: "Where can I find horse riding lessons in Toronto?" It returns a categorized list of local stables.
• 4:30 Tracing Agent's Thought Process: The video shows the most interesting feature: the "Events" tab in the ADK web UI. This interface displays Gemini's "thought signature" and reveals the series of increasingly specific web searches it performed (e.g., from generic searches to checking for public transit access and operational status) to generate the final answer.