Sign in to continue learning
Building Live Voice Agents with Google’s ADK
Intermediate
1 hour 15 mins
Topics
Agents
In this lesson, you learn how to get back a structured response from a region using a defined schema and see the results as an external file. We also see how we can change the behavior of a main agent as a coordinator rather than a newsreader. All right. Let's go. In our last lesson, we built a fantastic interactive news assistant. It could hold a conversation, fetch news and financial data in real time, and discuss the headlines with us. But now we're going to evolve that concept into much more powerful pattern, which is called as the coordinator dispatcher model. Think about it for real world tasks like preparing the search for a podcast. You don't want to sit and listen to any and read out every single headline it finds. That's inefficient. Instead, you want to act like a director. You give a command and your research assistant goes off, does the work quietly in the background, and then comes back to you. Only when it's complete formatted report is ready. That's exactly what we are building in this lesson. are transforming our agent from the conversationalist into a silent, efficient background researcher. It will take our request coordinated tool to perform the research and dispatch the final output to a file, only updating us on its progress. To achieve. This will be making three key changes. First, we are adding a new tool to its toolkit. Second, we are fundamentally rewriting the instruction for a root agent to change its behavior. and third, the way we interact with it will be much simpler focused on initiating the task and receiving the final product. As with the previous lessons, the initial setup is the same. We first ensure the latest version of VDC is installed and all the other dependencies are also there. Then we create the folder and boilerplate files for our agent. Using a k create. Since we have done this few times, let's jump straight to the code changes. Next let's start with the tools. Just as in lesson three are agents still has the get financial context function. Jobs remains the same. To take a list of stock tickers and return the current price and daily percentage change. We do not need to change anything. There. Now, the new addition to our agent dot profile is a function called as save news to markdown. This is another custom function to it's job is very straightforward. It takes two arguments file name and content to be saved. Inside the function, it simply writes the content into a markdown file. The tool is the final step in our agent workflow, allowing it to save its complete research to an external file, which we named as a research report.md. Now, for the most important part, the new instructions of a root agent. This is where we define its new personality as a background AI research coordinator. The instructions are now incredibly specific to enforce its new behavior. First, we have implemented a strict two message interaction workflow. When you make a request like find me the latest news, the agent's only immediate response will be an acknowledgment like, okay, I'll start researching the latest news. This might take a moment. After sending that message, the agent goes silent. This is the background processing phase. It is now executing a precise sequence of dual calls. What we defined in its instruction in. First, use Google search to find five articles. Then it will internally extract the stock tickers from those articles. Next it will pass those tickers to our Get financial tool. To ensure the output is perfect, every time we have embedded a required report schema directly into the agent's instruction. This is one of the most powerful feature of using large language model like Gemini with the SDK, by providing a clear markdown template in the prompt. We're giving the model a precise blueprint to follow. It understands the structure and will meticulously format the research to match it exactly. It is worth noting for more complex data structures, the SDK allows you to define a pedantic schema and pass it as an output schema. We'll explore that more advanced pattern when we get to multi-agent systems. but for our current needs, defining the schema directly in the prompt is a highly effective and simple way to guarantee a consistent output. Finally it calls are saved. News to markdown tool passing in the file name a research report.md and the formatted content. Only when the file is successfully saved. Does the agent speak for the second and final time. It will deliver its confirmation message. Like all done, I've compiled the research report and saved it to a researcher. Performed. This entire complex research process happens autonomously in the background. Between that first acknowledgment and the final confirmation. Now let's run it. You already know the drill. We will start the terminal. Navigate to our directory and run the web command to start the server. Once you open the web UI, I just give it a simple command. Get me the latest news about artificial intelligence. Okay, I'll start researching the latest AI news. I will enrich the findings with financial data and compile a report for you. This might take a moment. All done. I've compiled the research report with the latest financial context and saved it to AI. Underscore research. Underscore report deemed. All right, now let us look at the output. Remember, we said that this is a background agent, which means it took our instructions or requirements, and then it did the work in the back end and then save the file. Now let's look at the file that it has saved and see what exactly do we have and what it has done as the background processing. Wonderful. As you can see here, the report starts with a clear title, which is a industry news report followed by a top headline. Each of the five news item is neatly formatted with the headline itself, followed by a bullet point detailing the company and its stock ticker and the market data showing the current stock price and percentage change. And finally, it also shows you the brief summary of the article. Now this is consistent, predictable output is the direct result of providing that clear schema in our agents instruction, demonstrating how you can actually guide the model to produce exactly the structured data you need for your downstream task.
