Build AI Agent Application with Agent Development Kit (ADK) | Get Started with Google's Agent SDK

Google’s Agent Development Kit (ADK) is positioned as a practical way to build agentic applications with a clear workflow structure, built-in deployment options, and evaluation tooling—demonstrated through a working “diet coach” agent that turns a user’s ingredients and dietary goals into recipe recommendations with computed nutrition macros.

In the live demo, a web chat session (“chef’s kiss”) collects two inputs: the ingredients available in the kitchen and the dietary targets (example: high protein and low calories). The conversation then hands off to a coordinator agent that routes the request to specialized sub-agents. One agent extracts the user-provided ingredients and goals; another agent searches a local recipe set using a custom tool (“find recipes”), allowing up to two missing ingredients. The system returns a shortlist of matching recipes—such as “quasi grilled cheese,” “chicken cadidilia,” and “apple slices with peanut butter”—and then triggers a second custom tool (“calculate recipe macros”) for each candidate.

Those tool calls compute nutrition totals—calories, protein, carbs, and fat—by scaling ingredient-level macro data stored in JSON files. With the computed macros in hand, the recommender agent selects the best fit for the stated goals. In the example, “chicken cadidilia” stands out with roughly 50g of protein, aligning with the high-protein, low-calorie preference. The UI also exposes the internal execution trace: agent-to-agent transfers, tool invocations, and intermediate results, making it easier to debug what the agent is doing rather than treating it as a black box.

Beyond the demo, ADK is presented as more than a prompt-and-chat wrapper. The library supports agentic workflows with sequential, parallel, and looping execution patterns, and it encourages composing multiple specialized agents in a hierarchy. Google’s tooling also includes a Google Search tool (not used in the demo) and deployment pathways such as Docker deployment and an intended integration with Vertex AI. A key differentiator highlighted is built-in evaluation support: instead of relying on online demos, teams can create evaluation sets and inspect how agentic applications behave under realistic conditions.

The walkthrough then shifts from UI to code. The project uses UV for dependency management and shows how to install ADK (via pip install Google ADK). It loads two JSON “databases”: one with ingredient macros per 100g and another with recipe ingredient lists and gram weights. The agent logic is split into three roles: a user info agent (collects ingredients and goals), a recipe finder/recommender agent (calls tools to find recipes and compute macros), and a coordinator/root agent that orchestrates transfers. Tool functions are plain Python functions with typed arguments and docstrings that ADK parses for tool definitions. Running the app starts a local server (ADK web on port 8000), and the execution trace in the UI mirrors the underlying agent calls.

Finally, the tutorial notes practical compatibility issues: attempts to use other models via Ollama ran into tool-calling failures, while the flow worked with Gemini 2.0 Flash. Overall, ADK is framed as a structured, debuggable foundation for agentic apps—especially when paired with Gemini models—while still leaving open questions about how smoothly it supports non-Google/open-model backends at the time of release.