Google ADK
★ 20.4k · GitHub →Google's code-first agent toolkit: build, evaluate, and deploy agents with a full local dev UI.
Install
$ pip install google-adk What is Google ADK?
The Agent Development Kit (ADK) is Google's answer to the agent framework question, and it shows its origins: this is the same framework powering agents inside Google products, released as Apache-2.0 open source. It treats agent development like software engineering — agents are code, workflows are composable (sequential, parallel, loop, and since 2.0 a graph-based Workflow Runtime), and evaluation is a built-in step rather than an afterthought. The 2.0 release added a Task API for structured agent-to-agent delegation with human-in-the-loop support.
Two things distinguish it in practice. First, the tooling: `adk web` launches a local development UI where you chat with your agent, inspect every event and state change, and debug tool calls — the best out-of-box dev experience among the Python agent frameworks. Second, the deployment story: agents map directly onto Vertex AI Agent Engine for managed hosting, with Cloud Run as the DIY path, plus A2A protocol support for cross-vendor agent interop.
The trade-offs are what you would guess. ADK is model-agnostic on paper (LiteLLM adapters cover other providers) but Gemini-first in polish and documentation, and the framework surface is large — agents, tools, sessions, artifacts, callbacks, evaluators, workflows — with a release cadence that has included breaking changes between major versions. Teams outside Google Cloud get a capable framework; teams inside it get the full experience.
Key capabilities
- ▸ Code-first agent definitions with composable workflows: sequential, parallel, loop, and graph-based runtime
- ▸ `adk web` local development UI for interactive chat, event inspection, and step-through debugging
- ▸ Built-in evaluation framework for testing agent response quality and execution trajectories
- ▸ Direct deployment path to Vertex AI Agent Engine or Cloud Run, with A2A protocol interop
- ▸ Model-agnostic via adapters, with first-class Gemini integration and multi-agent hierarchies
When to use it
ADK is the obvious pick if you are on Google Cloud and Gemini: the dev UI, evaluation tooling, and Vertex AI deployment form a coherent pipeline nothing else matches inside that ecosystem. It also suits teams who want engineering discipline — evals, artifacts, structured workflows — baked into the framework. Skip it if you are not on GCP and just need a lightweight agent loop: the surface area is heavy for simple tasks, the Gemini-first bias is real, and Pydantic AI or the OpenAI Agents SDK will get you there with far less framework to carry.
Alternatives
Frequently asked questions
Does Google ADK only work with Gemini models?
No. ADK is model-agnostic through adapter integrations (including LiteLLM), so Claude, GPT, Llama, and others work. In practice, though, Gemini gets the first-class treatment: documentation examples, streaming features, and Vertex AI integration are built around it. If Gemini is not in your stack, evaluate whether the non-Gemini experience meets your needs before committing.
How do I deploy an ADK agent to production?
Three main paths: Vertex AI Agent Engine (Google's managed runtime for ADK agents — the lowest-effort option), Cloud Run (containerize and self-manage on GCP), or any container platform since ADK agents are ordinary Python apps. The framework also supports the A2A protocol, letting deployed agents be discovered and called by agents from other vendors.
What is the difference between Google ADK and Vertex AI Agent Builder?
ADK is the open-source, code-first framework — you write Python, run locally with `adk web`, and control everything. Vertex AI Agent Builder is the umbrella for Google Cloud's managed agent products, including no-code/low-code options. They connect: ADK-built agents deploy onto Vertex AI Agent Engine, so ADK is the pro-code entry into the same platform.
Star count updated 2026-07-02 · refreshed weekly from the GitHub API · browse all tools →