Google ADK: Agent Development Kit Explained

Google ADK, the Agent Development Kit, is an open source framework from Google for building, evaluating and deploying AI agents. An agent is written as code: you declare the model, the instructions, the tools it may call and the other agents it can delegate to, then you run it locally, test it and ship it to the cloud.

Three ideas set ADK apart from a plain LLM wrapper. Agents compose: a coordinator can route requests to specialists, and workflow agents run steps in sequence, in parallel or in a loop, so a multi-agent system fits in a few dozen lines. Tools are first class: plain functions, built-in tools such as Google Search and code execution, MCP servers and even other agents can be exposed to the model, with the framework handling the call loop. The lifecycle is covered end to end: a local developer UI to inspect every step, an evaluation harness to score runs against expected behaviour, and a deployment path to Cloud Run or Vertex AI Agent Engine. ADK was designed around Gemini but connects to other models, and it speaks the A2A protocol so agents built with other stacks can collaborate with it.

In my projects I reach for ADK when the deliverable is a real multi-agent system on Google Cloud, for example a sales assistant where one agent qualifies the lead, another checks the calendar and a third drafts the proposal. When the need is one AI feature inside a Next.js or Firebase app, I stay on Genkit, which is lighter and built around typed flows. The two combine well: a Genkit flow can call an ADK agent as a single step.