Google ADK (Agent Development Kit)

Google ADK (Agent Development Kit) is Google's open-source framework for building, orchestrating, and deploying AI agents. ADK provides a modular, code-first approach to agent development with built-in support for multi-agent architectures, tool integration, and evaluation.

ADK integrates natively with Google's Gemini models and Vertex AI platform while remaining model-agnostic — developers can use any LLM as the agent's reasoning engine. The framework supports complex agent topologies including sequential pipelines, parallel execution, and hierarchical delegation between specialized sub-agents.

ADK 2.0 and the Workflow Runtime

As of October 2026 the current major version is ADK 2.0. Google's documentation lists general availability of v2.0.0 for Python on May 19, 2026, for Go on June 30, 2026, and for TypeScript on August 21, 2026. ADK also ships Java and Kotlin SDKs, but the 2.0 documentation gives release dates only for those three languages. The project's documentation has moved from google.github.io/adk-docs to adk.dev, which describes ADK as "the open-source agent development framework that lets you build, debug, and deploy reliable AI agents at enterprise scale."

The change in 2.0 is architectural. ADK 1.x ran agents as a hierarchy of executors. ADK 2.0 introduces a Workflow Runtime in which, in the documentation's words, "Agents, Tools, and Functions are evaluated as individual nodes within a workflow graph." Three styles sit on top of it: graph-based workflows for deterministic routing, dynamic workflows that use code for loops and branching, and collaborative workflows built from a coordinator agent and subagents, which is ADK's version of the orchestrator–worker pattern. A workflow can pause for human input and resume, and the framework retries failed steps automatically.

Google describes 2.0 as designed to be compatible with 1.x agents, with breaking changes. Events gain new fields that track graph state, so a 2.0 event written to a rigid 1.x database table fails. Agents are now nodes: in Python, BaseAgent subclasses BaseNode, and legacy overrides such as _run_async_impl() are silently ignored. Events must be yielded from the node, not appended to the session list by hand. The 1.x lines remain installable for teams that are not ready to migrate.

A2A, MCP and Models

ADK treats the two main agent protocols as complements. For tools, its McpToolset class connects an agent to any Model Context Protocol server over local stdio or remote HTTP, and ADK tools can be wrapped the other way and served as an MCP server. For agents, ADK can expose an agent over Google's A2A protocol or consume a remote A2A agent alongside local subagents; the documentation still marks its A2A support as experimental. Beyond Gemini, the documentation covers Gemma, Claude and OpenAI models, and local or self-hosted models through Ollama, vLLM and LiteLLM.

Deployment and Evaluation

The documentation describes four deployment routes: Agent Runtime on Google's Agent Platform, a fully managed auto-scaling service for agents; Cloud Run, for running an agent as a container; Google Kubernetes Engine; and any other container-friendly infrastructure, including offline systems. ADK also includes tooling for evaluation, session management with rewind, and context compression, which makes it a harness as well as a library. Whether a graph of several agents beats one well-equipped agent on a given task is a separate question that agent evals have to answer, and the 2026 research on multi-agent systems suggests the answer is often no.

Google ADK competes with LangChain, CrewAI, and OpenAI's Agents SDK in the agent orchestration layer, offering Google's perspective on how multi-agent systems should be structured and deployed at scale.