Building AI Agents with ADK: Unlocking Stateful Workflows

Artificial intelligence development is changing fast. Instead of using simple, one-off interactions with AI models (stateless request-response cycles), developers now focus on building complex, multi-turn agentic workflows. In these new workflows, AI agents can remember past interactions and hold longer, more meaningful conversations. As a result, this helps them complete complex tasks more effectively.

Takeaways 🚀
✅Shift to stateful AI workflows
✅Unified access to AI models
✅Simplified conversation state management
✅Transparent A2A protocol integration
✅Enhanced agentic AI capabilities

To keep up with these changes, Google recently launched the Interactions API in beta. With this update, developers now have one unified tool that offers a single interface for accessing both raw AI models and the fully managed Gemini Deep Research Agent. Moreover, if you already use the Agent Development Kit (ADK), this API unlocks exciting new ways to build and connect advanced AI agents.

In this post, you’ll see exactly where the Interactions API fits in today’s developer toolkit. Specifically, you’ll learn two main ways to use this technology: powering your ADK agents, and then connecting with other built-in agents through a simple, transparent bridge.

A New Era of AI: Stateful Workflows

In the past, interacting with AI meant sending a request and getting a response. However, the AI forgot everything from earlier conversations—each time, you started fresh. While this stateless approach worked for simple questions, it made it hard for AI to manage complex, multi-step tasks.

Stateful workflows, on the other hand, change everything. With stateful models, an AI agent keeps track of the entire conversation from start to finish. As a result, it remembers past topics, tracks what actions it took, and always stays focused on your main goal. This memory lets AI agents work as researchers, planners, or assistants that manage tasks needing several steps and on-the-fly changes. Because of these capabilities, the Interactions API was built specifically for stateful workflows. It helps developers easily guide complex conversations and, consequently, manage detailed tasks more effectively.

Pattern 1: Powering Your ADK Agents with the Interactions API

When you build an agent with the Agent Development Kit (ADK), you use a large language model (LLM) like Gemini to generate thoughts, make plans, call tools, and create responses. Previously, you did this through an endpoint called generateContent.

Now, the new Interactions API gives you a faster and easier way to manage the complex state that agentic behavior needs. When you upgrade your agents to this endpoint, you gain several important benefits that are built for today’s AI workflows.

Why Make the Switch?

• Unified Access to Models and Agents: The same API endpoint can be used to access a standard model like Gemini 3 Pro or a sophisticated built-in agent like the Gemini Deep Research Agent. This simplifies the development process.
• Simplified State Management: The API allows you to offload the management of conversation history to the server. This reduces the amount of boilerplate code needed in your ADK agent, letting you focus on the unique logic of your application.
• Support for Background Execution: Many complex tasks can take a long time to complete. The Interactions API supports long-running tasks through a background execution mode. The API can return an interaction ID immediately, allowing the reasoning loop to run on the server. Your application can then check the status asynchronously to get the final result without timing out.
• Native Handling of “Thoughts”: The API explicitly separates an agent’s internal “thoughts” or reasoning steps from its final response. This makes it easier for your ADK agent to process and understand the agent’s reasoning chain.

With this pattern, you control the main logic and decisions in your ADK code. Meanwhile, the Interactions API handles the heavy lifting—like tracking context and managing the agent’s state. As a result, you get clear control over both your agent’s own tasks and the detailed reasoning happening inside the API. Ultimately, this separation of concerns makes your workflow more efficient and easier to maintain.

Pattern 2: A Transparent Bridge for A2A Agents

The Agent2Agent (A2A) protocol lets different AI agents talk to and work with each other. If you already use A2A agents, you might want them to access powerful tools like the Gemini Deep Research Agent. Previously, you had to write a custom adapter to connect a new third-party API.

Now, with the latest advancements, this process becomes much simpler.

The new InteractionsApiTransport makes this process much easier. Specifically, this tool connects the A2A protocol right to the Interactions API, so the API can “speak” A2A directly. As a result, you can use any Interactions API endpoint as a remote A2A agent. Even better, there’s no need to change your existing clients—they simply see another agent they can talk to as usual.

How the Bridge Functions

To make everything work smoothly, the InteractionsApiTransport layer quickly converts commands between both protocols:

• An A2A SendMessage command turns into an Interactions API create call.
• An A2A Task matches to an Interaction ID.
• The TaskStatus in A2A lines up with the Interaction Status.

In addition, with this approach, the Interactions API feels invisible in your workflow. Consequently, you can use new tools like the Gemini Deep Research Agent right away, without needing to rewrite your multi-agent system. Your A2A client code stays the same, and the transport also handles streaming updates from the agent in real time.

• An A2A SendMessage command becomes an Interactions API create call.
• An A2A Task is mapped to an Interaction ID.
• The TaskStatus in A2A corresponds to the Interaction Status.

Therefore, this approach keeps the Interactions API invisible for developers. You can use new tools like the Gemini Deep Research Agent right away—no need to rewrite your multi-agent system. Additionally, your A2A client code stays the same. Furthermore, the transport also manages streaming events, so you get real-time updates from the agent.

The Future of Agentic AI

With these innovations, the Gemini Interactions API completely changes how you build and use AI. As a result, you now have the tools to move beyond basic question-and-answer bots. Instead, you can create smart agents that handle long-running, stateful tasks and consistently remember what happened during each conversation.

For example, you can use this new API whether you build custom agents with the ADK or connect a group of agents with the A2A protocol. In both cases, it works as a strong inference engine and a flexible remote agent. This versatility lets you grow your AI system’s abilities with little hassle. As a result, with these tools in hand, you can create smarter, more capable, and more helpful AI applications for everyone.