Welcome to Building Live Voice Agents with Google's ADK, built in partnership with Google and taught by Lavi Nigam and Sita Lakshmi Sangameswaran, who are both machine learning engineers at Google Cloud AI. In this course, you learn how to build multi-agent AI applications using Google's Agent Development Kit, or ADK. You build a podcast AI agents that gets live voice input from the user, researches the topic, plans an episode, drafts a multi-person transcript and finally produces a podcast audio file. Google's ADK is an open source framework that makes it easy to build simple agents with just a few lines of code and can scale up to complex multi-agent systems. ADK includes built-in graph flow patterns to control the behavior of different agents and also supports flexible LLM driven orchestration, where an LLM determines what sequence of steps to take. It provides common agent building blocks, including models, tools, memory, and observability features like providing traces and evals. This course also acts as a quick introduction to live voice conversational interfaces with agents, which you'll be able to build with just a few lines of code. We'll also explore multi-agent systems where each agent is focused on a specific task, and multiple of such agents collaborate to accomplish a more complex goal. At the end, you have built your own agentic system to generate podcasts. Thanks Andrew. To start this course, you'll create a simple voice agent using Google ADK and Gemini Live model. Your agent will take in user voice input, use a language model to reason and then send back a voice output. Then you'll learn about the core building blocks of ADK, like sessions, state, and memory. Session is like a container that keeps all parts of one interaction together. If you start talking to an agent, the session begins and everything you say and the agent replies stays inside that session. Similarly, state is the scratchpad or the short-term memory for your agent. It tracks the agent progress so that the agent doesn't lose track of what's doing in the mid-flow. Next is memory. Memory extends this further by letting the agent recall past inputs, outputs, or tool calls and use when responding. Together, these primitives form the foundation of every agent you'll build with ADK. Agents become much more useful with tools, such as Google search or any external API. This gives your agent the ability to access a wider world, fresh and factual information and to take actions. Tools can be functions in your code or MCP service or public APIs. As your agentic system becomes more capable, it can do more things and you need to constrain or control what it can and what it should not do. This is easily done through callbacks in ADK, which lets you intercept before and after every single model tool and agent call. You can add code or remote APIs or simply a specialized agent to handle each of these callbacks. You start by building a single podcast agent and then turn that into a multi-agent system. In the multi-agent system, you can have one agent carry out or figure out the plan, and then a second one to do background research, and a third one to write. And you'll see how to decompose a complex task into subtasks that different agents can carry out, which is, turns out, an important design pattern for implementing agentic workflows. Many people have contributed to this course. I'd like to thank from Google, Bo Yang, Hangfei Lin, Alan Blount, Julia Wiesinger, Polong Lin, Erwin Huizenga, and many others behind the scenes who have made ADK possible. From DeepLearning.AI, Brendan Brown and Esmaeil Gargari also contributed to this course. This course will help you learn not only agent building with Google ADK, but also how to bring voice capabilities to your AI applications. In the next lesson, you'll build your first voice agent. So, let's go to the next video to get started. Agents become much more useful with tools such as Google Search or any external API, and this gives your agent the ability to access a wider world. But to get fresh and factual information and to take actions. Tools can be functions in your code or Ncbi service or public API.