This website uses cookies

Read our Privacy policy and Terms of use for more information.

Last Updated: November 18, 2026

Summary: Google's Agent Development Kit (ADK) is a free, open-source framework for building, testing and deploying AI agents in code, with official quickstarts for Python, TypeScript, Go, Java and Kotlin. ADK is optimized for Gemini but can work with other models, supports MCP tools and the A2A protocol, and deploys to Agent Runtime on Gemini Enterprise Agent Platform, Cloud Run or Google Kubernetes Engine. The framework costs nothing; you pay for the models and cloud resources your agents use.

Google released ADK in April 2025 and has rebuilt much of it since: ADK 2.0 added a graph-based workflow engine, new languages arrived, and the Google Cloud services around it were renamed. Many tutorials still describe the older version. Details below were checked against official documentation in November 2026.

Free download: The AI Stats Sheet 2026. 36 key AI numbers for business on four printable pages, with the source under every figure. Join 2,000+ business leaders who read AI Business Weekly and get it in your inbox right away.

In this guide

What Is Google ADK?

Google ADK (Agent Development Kit) is an open-source framework for building AI agents in code. Google Cloud's ADK overview describes it as "an open-source agent development framework that lets you build, debug, and deploy reliable AI agents at enterprise scale."

Google announced ADK on April 9, 2025. The Google Developers Blog launch post said ADK is the same framework that powers agents inside Google products such as Agentspace and the Google Customer Engagement Suite.

The ADK Python README calls the kit model-agnostic and deployment-agnostic, though optimized for Gemini. Google ADK gives developers building blocks (agents, tools, sessions, memory, evaluation) plus a command line tool and a local web UI for debugging. The code uses the Apache 2.0 license.

New to agents? Start with our explainer on what AI agents are. Official ADK documentation now lives at adk.dev.

Which Languages Does Google ADK Support?

Official sources differ on the count. The ADK get started page on adk.dev lists quickstarts for five languages: Python, TypeScript, Go, Java and Kotlin. Google Cloud's ADK overview says ADK is available in four: Python, TypeScript, Go and Java. IT Brief reported on October 27, 2026 that Google supports five.

The languages are not all on the same release line. The ADK 2.0 page says version 2.0 covers Python, TypeScript and Go only.

Language

Current status

Key date

Python

ADK 2.x; check PyPI for the latest release

2.0 generally available May 19, 2026

Go

ADK 2.x (new module path google.golang.org/adk/v2)

2.0 generally available June 30, 2026

TypeScript

ADK 2.x

2.0 generally available August 21, 2026

Kotlin

ADK for Kotlin 1.0, including Android extensions

1.0 announced September 9, 2026

Java

Supported, with its own quickstart; not part of the 2.0 release

No 2.0 date announced

Python releases ship often, so check the google-adk package on PyPI for the current version. The Kotlin date comes from Google's ADK for Kotlin 1.0 announcement.

How Does Google ADK Work? The Core Concepts

Google ADK combines agents, tools and a runner that manages each conversation's state and history, according to the ADK technical overview.

Developer working at a desk with two computer monitors, writing code for an AI agent

ADK agents are defined in code: an agent, its tools and a runner that tracks each session. Photo: Behnam Norouzi / Unsplash+

Agents

An ADK agent is a unit built for a specific task. ADK has two broad kinds. An LlmAgent uses a language model to reason, decide and call tools. Workflow agents (SequentialAgent, ParallelAgent and LoopAgent) run steps in a fixed, predictable order without asking a model what to do next.

ADK tools

ADK tools let an agent call APIs, search or run code. According to the ADK README, tools can be pre-built tools, your own functions, OpenAPI specs, MCP tools or other agents, and a confirmation flow can require human approval before a tool runs.

Multi-agent systems and workflows

Specialized agents sit in a hierarchy, and a parent agent can delegate work either by handing control to a sub-agent (LLM-driven transfer) or by calling the sub-agent as a tool.

ADK 2.0 added the Workflow Runtime, which moves agent execution from a hierarchical executor to a graph-based engine. The ADK 2.0 page lists three patterns: deterministic graph workflows, dynamic workflows with code-driven loops and branching, and collaborative workflows where a coordinator agent directs sub-agents.

Sessions, state and memory

The ADK sessions documentation defines three layers:

  • Session: a single, ongoing interaction between a user and your agent system.

  • State: data stored within a specific session.

  • Memory: a store of information that can span many past sessions or include external data.

For local development, ADK offers in-memory services. For production, the ADK memory page lists Memory Bank, which it calls "a fully managed Google Cloud service" for persistent agent memory.

Evaluation

The ADK evaluation guide says it checks two things: the trajectory (which tools the agent called and in what order, compared with an expected path) and the quality of the final response.

Developers can run evaluations in the adk web interface, inside a pytest suite for CI/CD, or with the adk eval command. Built-in criteria include trajectory scores, rubric-based quality checks, hallucination and safety checks, plus token usage metrics.

Where Can You Deploy ADK Agents?

ADK agents can run locally or deploy to three Google Cloud targets: Agent Runtime, Cloud Run and Google Kubernetes Engine (GKE). Google Cloud's ADK overview says you can "run agents locally or scale them globally using Runtime, Cloud Run, or Google Kubernetes Engine." The ADK README also shows a Docker deployment.

Aisle between rows of server cabinets in a data center, where cloud-deployed AI agents run

Agent Runtime, Cloud Run and GKE are the three documented Google Cloud targets for ADK agents. Photo: Getty Images / Unsplash+

Target

What it is

Best for

Languages with a documented path

Agent Runtime

Managed agent hosting on Gemini Enterprise Agent Platform (the platform formerly called Vertex AI)

Teams that want Google to run, scale and govern agents, with managed sessions and Memory Bank

Python, Go

Cloud Run

Google's serverless container platform

A simple, scalable web service you control

Python, TypeScript, Go, Java

GKE

Managed Kubernetes

Teams already on Kubernetes that need full control of infrastructure

Python (automated command); others via manual setup

The ADK Agent Runtime guide describes Agent Runtime as "a set of modular services that help developers scale and govern agents in production." Teams can deploy through the ADK command line or use Agents CLI, which also sets up CI/CD pipelines. Older tutorials call this service Agent Engine: Google Cloud's Agent Platform release notes for April 22, 2026 state that "Agent Engine is now Agent Runtime."

The ADK Cloud Run guide uses adk deploy cloud_run for Python, an adkgo tool for Go and gcloud run deploy for Java. The ADK GKE guide notes that adk deploy gke is Python only.

For Google's wider agent lineup, see our guide to Google AI agents.

Does Google ADK Support A2A and MCP?

Yes. Google ADK supports the Model Context Protocol (MCP) for connecting agents to tools and data, and the Agent2Agent (A2A) protocol for agent-to-agent communication, though A2A is experimental in some languages.

Software team reviewing sticky notes on an office whiteboard while comparing agent framework options

MCP connects agents to tools and data, while A2A handles communication between agents. Photo: Getty Images / Unsplash+

MCP support

The ADK MCP tools page lists MCP support in all five languages. In Python, an McpToolset added to an agent's tools list connects to an MCP server and discovers its tools. ADK can also go the other way: its to_mcp_server function wraps ADK tools in an MCP server so other clients can use them.

A2A support

The ADK A2A documentation covers both exposing your agent so other agents can call it and consuming a remote agent from your own. Quickstarts exist for Python, Go and Java. The same page states that "A2A support is experimental in ADK Go, ADK Java, and ADK Kotlin," which leaves Python as the one language with an A2A quickstart not flagged as experimental.

Is Google ADK Free?

Yes, Google ADK itself is free. The framework is open source under the Apache 2.0 license, and you can run it on your own laptop at no cost. You pay for two things once an agent does real work: the model it calls and the infrastructure it runs on.

  • Model costs: ADK is optimized for Gemini, and the Python quickstart uses the gemini-flash-latest model with a Google API key. Our Gemini API pricing guide breaks down the per-token rates.

  • Hosting costs: Cloud Run and GKE bill at their normal Google Cloud rates. Agent Runtime has its own rates, listed on the Gemini Enterprise Agent Platform pricing page.

Agent Runtime resource

List price

1-year commit

3-year commit

Agent Compute (per vCPU-hour)

$0.085

$0.0765

$0.068

Agent Memory (per GiB-hour of RAM)

$0.009

$0.0081

$0.0072

Sessions and Memory Bank storage

$0.30 per GiB-month

n/a

n/a

Usage is rounded to the nearest second, and idle time between turns is not billed. Sessions and Memory Bank reads cost $0.085 per 3 million operations and writes $0.085 per 1 million, billed since September 1, 2026. New Google Cloud customers get $300 in free credits.

A worked example

Say an agent runs on Agent Runtime with 1 vCPU and 2 GiB of memory, and is actively working for 100 hours in a month. At list prices, compute costs 100 x $0.085 = $8.50 and memory costs 200 GiB-hours x $0.009 = $1.80, for $10.30 in hosting. Gemini token costs come on top.

Google ADK Tutorial: The Python Quickstart in Seven Steps

These steps come straight from the official ADK Python quickstart. The quickstart lists Python 3.10 or later; the newest README on GitHub lists Python 3.11 or later, so use 3.11+ to be safe.

  1. Create a virtual environment (recommended): python3 -m venv .venv

  2. Activate it on macOS or Linux: source .venv/bin/activate (Windows users run .venv\Scripts\activate.bat or .venv\Scripts\Activate.ps1)

  3. Install ADK: pip install google-adk

  4. Create a project: adk create my_agent

  5. Add your API key to my_agent/.env: echo 'GOOGLE_API_KEY="YOUR_API_KEY"' > .env

  6. Chat in the terminal: adk run my_agent

  7. Or open the web UI from the parent folder: adk web --port 8000, then visit http://localhost:8000

The quickstart warns that ADK Web is for development and debugging only, not production. For a framework-neutral walkthrough, read our guide on how to build AI agents.

Free download: The AI Stats Sheet 2026. 36 key AI numbers for business on four printable pages, with the source under every figure. Join 2,000+ business leaders who read AI Business Weekly and get it in your inbox right away.

Google ADK vs OpenAI Agents SDK, Claude Agent SDK, LangGraph and CrewAI

Here is how ADK and its rivals compare, based on their official documentation:

Framework

Languages

Models

Multi-agent approach

MCP / A2A

Managed hosting

Google ADK

Python, TypeScript, Go, Java, Kotlin

Optimized for Gemini; other models supported

Agent hierarchies plus a graph-based Workflow Runtime

MCP in all five languages; A2A (experimental in Go, Java, Kotlin)

Agent Runtime on Gemini Enterprise Agent Platform

OpenAI Agents SDK

Python, TypeScript

Provider-agnostic: OpenAI APIs plus 100+ other LLMs

Handoffs and agents as tools

MCP tools; A2A not covered in the docs cited here

Separate Agents API runs a managed harness in OpenAI's service

Claude Agent SDK

Python, TypeScript

Claude

Subagents for focused subtasks

MCP; A2A not covered in the docs cited here

Separate Claude Managed Agents product

LangGraph

Python, JavaScript

Works on its own or with LangChain's model integrations

Explicit graphs with durable, checkpointed state

Not covered in the overview docs cited here

LangSmith Agent Server

CrewAI

Python (3.10 to 3.13)

Not stated in the intro docs cited here

Flows that manage state, with Crews of collaborating agents

Not covered in the intro docs cited here

Not covered in the intro docs cited here

ADK vs OpenAI Agents SDK

The OpenAI Agents SDK README describes "a lightweight yet powerful framework for building multi-agent workflows" that is provider-agnostic. It centers on handoffs, guardrails, sessions and tracing. OpenAI's SDK guide says the SDK runs in your application, while its separate Agents API runs a managed harness in OpenAI's service. ADK covers more languages.

ADK vs Claude Agent SDK

The Claude Agent SDK overview says it gives developers "the same tools, agent loop, and context management that power Claude Code," programmable in Python and TypeScript. It includes built-in tools for files, commands and web search, plus hooks, subagents, MCP and permissions. Unlike ADK's Apache 2.0 license, its use is governed by Anthropic's Commercial Terms of Service. Our Claude AI agents guide covers Anthropic's options in depth.

ADK vs LangGraph

The LangGraph overview calls it "a low-level orchestration framework and runtime for long-running, stateful agents," with durable execution, human-in-the-loop pauses and memory. ADK 2.0's graph workflows now overlap with that idea. LangGraph's managed option is LangSmith Agent Server, while ADK's managed options run on Google Cloud.

ADK vs CrewAI

The CrewAI introduction describes "the leading open-source framework for orchestrating autonomous AI agents," built around Flows (which manage state and execution) and Crews (teams of agents that collaborate on tasks). It requires Python 3.10 to 3.13, per its installation guide.

Which agent development kit should you pick?

  • Pick Google ADK if you use Gemini or Google Cloud, need Go, Java or Kotlin, or want evaluation and managed deployment from one vendor.

  • Pick the OpenAI Agents SDK for a lightweight Python or TypeScript framework that works across many model providers.

  • Pick the Claude Agent SDK for agents that read, edit and run code or files, built on Claude.

  • Pick LangGraph for fine-grained control of state across any cloud.

  • Pick CrewAI for role-based agent teams in Python.

Our roundup of AI agent platforms compares no-code and enterprise options for teams that would rather not write code.

Frequently Asked Questions

What is Google ADK?
Google ADK (Agent Development Kit) is Google's open-source framework for building, evaluating and deploying AI agents in code. Its adk.dev docs list quickstarts for Python, TypeScript, Go, Java and Kotlin. It is optimized for Gemini models and released under the Apache 2.0 license.

What does ADK stand for?
In AI, ADK stands for Agent Development Kit, the name of Google's framework for building AI agents. Google uses the short form in its package and command names, such as google-adk and adk web.

Is Google ADK free to use?
Yes. The Agent Development Kit is free and open source. You pay only for the models your agent calls, such as Gemini, and for hosting, such as Agent Runtime at $0.085 per vCPU-hour and $0.009 per GiB-hour at list price.

Does Google ADK only work with Gemini?
No. Google says ADK is optimized for Gemini but designed to integrate with other language models. The official quickstart uses a Gemini model by default.

Where can I deploy a Google ADK agent?
Google documents three cloud targets: Agent Runtime on Gemini Enterprise Agent Platform (called Agent Engine in older docs), Cloud Run and Google Kubernetes Engine. ADK agents can also run locally or in any Docker container.

The Bottom Line

Google's Agent Development Kit has grown from a Python library into a multi-language framework with graph-based workflows, built-in evaluation, MCP and A2A support. The kit itself is free; models and hosting are where costs appear.

ADK makes the most sense for teams on Google Cloud or Gemini. Teams committed to OpenAI or Claude models should compare the vendor SDKs and LangGraph first.

Related Articles

⭐ Get more AI Business Weekly in Google: add us as a preferred source on Google.

Written by Sameer Khan with AI assistance, fact-checked against the sources linked above.