Build Intelligent Agents

1. Build Intelligent Agents

Hey, agent, can you help me plan my trip to San Francisco? OK, I can help with that. What's your budget for this trip? OK, let me pause it there and tell you that I just interacted with a multimodal multi-agent application authored in under 100 lines of code with a new agent development kit. Let's get into it. Hi, I'm Anand Dayar, a product manager here at Google Cloud. And I'm really excited to share something new with you today. We talked to a lot of AI developers and heard consistently that building sophisticated agents, particularly multi-agent or multimodal ones, is often way more complex and time consuming than it should be. We felt that friction firsthand building agents internally for our customers. And that's exactly why today we're thrilled to release a brand new open source project, the Agent Development Kit, or ADK for short. It is built directly from Google's experience creating production grade agents. Our goal with ADK is simple, to give developers a powerful open foundation for building complex production ready AI agents intuitively and with greater control. To us, a truly open foundation means it is model agnostic, it's deployment agnostic, and built for interoperability. So starting with model agnostic. Look, we obviously love our Gemini models, and especially the latest Gemini 2.5 Pro model, which has amazing reasoning capabilities. But with ADK, you can choose any model you want to use, whether it's a model from Google or anywhere else. ADK is also deployment agnostic. It can run locally wherever you do development, or on Google Cloud, or wherever your infra happens to live. It is also built for interoperability, so you can easily integrate your agents with your existing tools and services, or even integrate with agents that you may have built on other frameworks. Now, Bo, you led the engineering effort here. Why don't you share some of the design thinking behind ADK? Sure. Thanks, Anand. Hi, I'm Bo Yan, tech lead for agent development kit. One of our core principles behind this framework was to make agent development feel like software development. We deliberately built ADK using familiar software engineering patterns, just like regular classes or functions. Can I just say thank you? Seriously, Bo. As a developer that wants to build agentic application for production, that's just great to hear. Of course. Also, looking ahead, we saw huge potential for agents that interact naturally, agents that can hear, can see, can respond in real time. That's why native bidirectional audio and video streaming is fundamental, and it's baked into the core SDK. This unlocks the ability to create truly interactive and real-time agent experiences. Another common friction that we often hear is that it requires a lot of setup to just test and debug agent locally. Developers wanted to iterate quickly without spinning up a ton of complex dependencies. So we included a UI playground right in SDK. Now, with a single command, ADK web, you can launch a UI on your laptop to test, visualize, and debug your agent. By the way, you can be up and running in seconds, focusing purely on your agent's technology. The agent development kit is available today as a Python SDK with more languages like Java coming soon. With this ADK, you can build sophisticated multi-agent applications in just under 100 lines of code. OK, Ivan, do you want to show us a demo? Absolutely. So let's do this. So Anand and Bo talked about simplifying complexity. So let's build a three-planar agent super quickly. It's a multi-agent setup. We will have a main three-planar agent orchestrating two subagents, an idea agent to brainstorm a fun trip destination, and a refine agent to check if those ideas fit with a given budget. OK, so let's dive into the code. Here, I have my two agent definitions. And it's just standard ADK agent classes. So we give them an LLM, some instructions, and more importantly, clear description. So the parent agent understands their capabilities. So now we have our definition. They look good. They're super concise. So let's learn locally. Like Bo mentioned, it's just one command in the terminal, ADK web. So let's try it. And boom. OK, our local UI is now ready. So let's give our planner agent a task. Suggest a fun weekend from San Francisco under $200. Let's see. It's a tricky question. It's going to be a challenging task. Let's see. And there we go. We got a response. So now we have our agent working. And if you are a developer, you want to know what's happening under the hood. The Event tab shows the agent's workflow and how it decided to use the subagent. Here, you can see that the parent agent delegated to gather ideas and do a budget check. The Request and Response tab give us the step-by-step trace. So you know what data goes where. Yes, that's super useful. Having this tab that allows me to know what the action agent takes, what kind of function call he had, it's very useful. And it's actually essential for debugging, especially with multi-agents when you have multi-agents interacting between each other. So having this level of visibility locally is just fantastic. So now, Bo also mentioned that we have these features called native bidirectional audio and video streaming, which essentially means that agent can hear and see your input. It's built-in, no extra libraries needed for the basic setup. So let's try the audio. Anand, why don't you ask a question? Absolutely. Yeah, let's take it for a spin. Can you suggest an itinerary for two days in Seattle? I can definitely help you plan a two-day trip to Seattle. What kind of activities do you enjoy? What's impressive about this is how quickly you spun up this multi-agent application to have a live audio conversation. And I don't think you had to write any special code to get all this audio streaming working. No, that's right. ADK handles all the underlying streaming mechanisms for me behind the scenes. And as a developer, this speed up massively my development. It let me just focus on the agent logic, which in this case is planning the trip. And the core agent definition is still only 30 lines of codes for those three agents combined. We define them, we spin up the debug UI, and we had a live audio conversation all running locally. And that's just a glimpse of what's possible. With Agent Development Kit, you have the simplicity, control, flexibility, and that multimodal streaming capability native to the SDK. And what we've built with Agent Development Kit goes beyond just an SDK. It's a robust, open foundation built directly on Google's production experience. We're investing in the ecosystem to empower developers like you to build complex, powerful agentic solutions. So if you want to dive in, we've got the links right below in the description. And check out the ADK documentation for guides and concepts. Or if you're just like me and you prefer to jump straight into the code, head over to the GitHub repo. We included plenty of sample agents there to help you to get started with it. Great. Thank you for tuning in. And we're excited to see what you'll build with the Agent Development Kit.

2. Let's practice!

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.