Google ADK: Google's Agent Development Kit

Why Google built ADK

Google made the Agent Development Kit public in spring 2025, starting with Python. The docs sum up the ambition in one line: build production agents, not prototypes. The project is open source (Apache 2.0) and split into one repository per language: Python, TypeScript, Go and Java, with Kotlin also covered in the documentation.

An ADK AI agent combines a model, instructions and tools. What sets the kit apart is the range of ways to organize several agents: sequential, parallel or looping agents for predictable flows, and model-driven routing when the task is more open-ended. The result is a genuine multi-agent system with explicit orchestration.

What’s in the box

• a web-based dev UI to chat with your agent, follow every step and test it;
• an agent evaluation framework (criteria, user simulation, custom metrics);
• MCP tools and support for the A2A protocol so agents can talk to each other;
• deployment options: Agent Runtime (Google Cloud’s managed service), Cloud Run, GKE or any container platform.

Getting started

In Python you need version 3.11 or later. Install the google-adk package with pip, create an agent folder with its instructions and tools, then launch the dev UI locally. Equivalent packages exist for npm (TypeScript), Go and Maven/Gradle (Java, Kotlin). A Gemini API key is enough for your first experiments.

Strengths and limits

Strengths: many languages, varied multi-agent patterns, built-in evaluation, and a direct route to production on Google Cloud.

Limits: the ecosystem clearly leans toward Gemini and Google Cloud, even though other models work. The kit’s breadth means a steeper learning curve than a minimalist SDK, and the language versions don’t always move at the same pace. Hosting and model costs add up too, so estimate them with the AI agent cost calculator.

Alternatives

The OpenAI Agents SDK is lighter, LangGraph gives very fine control through a graph, and Microsoft Agent Framework fills a similar role in the Azure world. Within Google’s own stack, Gemini Enterprise Agent Platform (formerly Vertex AI Agent Builder) offers a more platform-integrated approach.