ADK

ADK stands for Agent Development Kit, Google's framework for building AI agents and multi-agent systems. It is designed to make agent development feel more like software development: define agents as structured components, connect them to tools, manage state and sessions, run them through a runtime, and deploy or evaluate them with built-in developer tooling.

How ADK actually works

ADK is built around a few core primitives. An agent is the execution unit. A tool gives the agent abilities beyond text generation. A session stores the state and event history for a run. A runner manages execution flow and coordinates what happens next. In Google's own docs, ADK supports both LLM-driven agents and more deterministic workflow-style agents, including sequential, parallel, and loop-based orchestration.

That distinction matters. ADK is not just a wrapper around a model call. It is a runtime for systems where one agent may use tools, delegate to another agent, keep track of state, stream output, and continue until the task is complete.

What it is good for

ADK is especially useful when you want one or more of the following:

• multi-agent systems with delegation
• native tool use across cloud services and external APIs
• stateful sessions and memory-like behavior during runs
• multimodal or streaming agents
• cloud-native deployment on Google's ecosystem

It is often the right choice for teams already building around Gemini, Google Search grounding, or Vertex AI-style infrastructure, because the framework is designed to fit that environment naturally instead of being bolted onto it later.

Initialization and first steps

For Python, Google's quickstart currently begins with a standard virtual environment and package install:

• python -m venv .venv
• source .venv/bin/activate on macOS/Linux or the equivalent on Windows
• pip install google-adk
• adk create my_agent

That adk create step scaffolds a starter project so you are not wiring the entire runtime manually from zero. From there, the usual next step is defining an agent in the generated project and running it locally through the ADK CLI or developer UI.

Where the docs actually are

Google's official documentation lives at https://google.github.io/adk-docs/.

The most useful starting points are:

• the overview: https://google.github.io/adk-docs/
• the technical overview: https://google.github.io/adk-docs/get-started/about/
• the Python quickstart: https://google.github.io/adk-docs/get-started/python/
• the multi-tool quickstart: https://google.github.io/adk-docs/get-started/quickstart/
• the Python API reference: https://google.github.io/adk-docs/api-reference/python/

Tradeoffs

ADK is powerful, but it is not the lightest starting point if all you need is one simple tool-calling loop. It becomes more valuable as your system gets more structured: multiple agents, explicit workflows, state management, evaluation, or deployment concerns. In other words, ADK shines when you are building an agent system, not just calling a model with a tool attached.