Google’s Agent Development Kit Now Covers Five Languages: Python, Go, TypeScript, Java and Kotlin

CONFIRMED: Google Cloud Blog, 13 October 2026.

Google says its Agent Development Kit (ADK), a toolkit for building AI agents, now has “coverage for Python, Go, TypeScript, Java, and Kotlin” [1].

An AI agent is software that uses an AI model to carry out tasks, often by calling other tools. Google says many agent frameworks “assume a Python-only stack”, which forces teams either to run a separate Python service or to build agent logic from scratch in their own language [1].

According to Google, each version of ADK is “built from the ground up for its own runtime and type system” but shares one architecture [1].

One set of protocols

Google says all five versions support the same open protocols [1]:

• MCP. “ADK includes a Model Context Protocol client in every language”, so a tool offered through an MCP server can be called from any of them.
• A2A. ADK supports the Agent2Agent protocol, “allowing an agent to communicate with a remote agent over HTTP”. Google says this lets agents written in different languages or frameworks work together.

Getting started

Google says developers can run a one-time setup command, uvx google-agents-cli setup, which checks Google Cloud sign-in, installs the command-line tool and adds ADK skills to coding agents such as Antigravity, Claude Code or Codex [1].

A scaffold command then builds a working project with A2A endpoints and deployment settings for Cloud Run, Gemini Enterprise Agent Runtime or Google Kubernetes Engine (GKE) [1].

What each language is for, according to Google

• Python is “ADK’s reference implementation”, with the most Google Cloud integrations and model adapters, including for Gemini, Anthropic, OpenAI and LiteLLM [1].
• Go is pitched at production back-end services. Google says it was “the first language after Python to ship the ADK 2.0 Graph Workflow runtime”, and that a compiled Go agent fits on “a roughly 2 MiB” base image [1].
• TypeScript uses the Zod library to define tools and ships “a standalone browser bundle”, so an agent can run inside a browser tab [1].
• Java fits teams using Maven, Gradle and Spring Boot, with bridges for Spring AI and LangChain4j [1].
• Kotlin checks tool definitions when the code is built, uses “zero runtime reflection”, and Google says it can “run the entire agent loop locally on Android” [1].

The blog post does not say when each language version first became available.

What this means

• Teams can keep their own language. Google’s pitch is that a company does not need to add Python just to build agents.
• Agents can be mixed. Because every version speaks MCP and A2A, Google says different parts of one system can be written in different languages.
• It ties into Google Cloud. The setup and scaffold tools point at Google’s own hosting options: Cloud Run, Gemini Enterprise Agent Runtime and GKE.

Sources

1. Google Cloud Blog (Karl Weinmeister), “One Agent Development Kit, five languages”, 13 October 2026. https://cloud.google.com/blog/topics/developers-practitioners/one-agent-development-kit-five-languages/

Related stories

• DXC Partners With LangChain and Licenses LangSmith for Its OASIS IT-Operations Agents
• LightOn Releases ColGrep-Agent-2B, an Open Local Code-Search Agent That It Says Cuts Claude Code’s Tokens by a Third
• Subquadratic Launches SubQ Snap, a Long-Context Decision Model for Enterprise AI Agents