Google ADK
Also known as: Agent Development Kit, ADK, Google Agent Development Kit
- Google ADK
- An open-source framework from Google for building, evaluating, and deploying AI agents and multi-agent systems with code-first control. Available in Python and TypeScript, it pairs with Google Cloud’s Gemini Enterprise Agent Platform for managed production deployment.
Google’s Agent Development Kit (ADK) is an open-source, code-first framework for building, evaluating, and deploying AI agents and multi-agent systems in Python or TypeScript.
What It Is
Most product teams meet ADK when they’re already shopping for an agent framework. They’ve looked at LangGraph for graph-style orchestration and CrewAI for role-based crews, and they want to know what Google brings to the same problem. ADK is Google’s answer: an open-source toolkit you install like any library, write agents in code, and run locally before deciding whether to deploy them on Google Cloud’s managed agent service.
ADK exists because building a single AI agent that calls a tool is straightforward; building a system where several agents coordinate, hand off work, and stay debuggable is not. It gives you primitives for the parts that get messy: defining an agent with its model, instructions, and tools; composing multiple agents into a hierarchy or workflow; persisting state between turns; running evaluations against expected outputs; and wiring in observability so you can see what each agent decided and why.
The framework is code-first, meaning you describe agents and their relationships in Python or TypeScript files rather than in a visual builder or YAML config. According to Google ADK Docs, the framework supports both languages, and according to Google Developers Blog, the TypeScript SDK is a more recent addition aimed at frontend and full-stack teams. According to Google Developers Blog, ADK was originally launched at Google Cloud NEXT 2025 and is positioned as a flexible toolkit for production multi-agent systems with precise control. The same code that runs on your laptop can be deployed to Google Cloud’s Gemini Enterprise Agent Platform — documented under Google Cloud Docs — which adds managed runtime, scaling, and identity integration on top of the open framework.
How It’s Used in Practice
The mainstream entry point looks like this: a developer installs ADK, defines a supervisor agent with a Gemini model and a few tools, then attaches sub-agents — each with its own narrow job, like searching docs, drafting an email, or querying a database. The supervisor decides which sub-agent handles each user turn. You run it locally, watch the trace, tweak the instructions, and only think about deployment once the agent behaves the way you want.
That’s the same shape as a LangGraph or CrewAI build, just inside Google’s stack. Teams typically pick ADK when they’re already on Google Cloud, when they want first-party Gemini integration without an adapter layer, or when they expect to graduate from a prototype to the managed Agent Platform without rewriting orchestration code.
Pro Tip: Build and test agents locally before reaching for the cloud runtime. The local CLI gives you fast feedback loops, and the same code deploys upward — there’s no benefit to starting in the managed environment, and it slows iteration when you’re still tuning instructions.
When to Use / When Not
| Scenario | Use | Avoid |
|---|---|---|
| You’re already on Google Cloud and want first-party Gemini integration | ✅ | |
| You need a vendor-neutral framework that runs identically across AWS, Azure, and GCP | ❌ | |
| You want to prototype locally in Python or TypeScript, then deploy a managed runtime | ✅ | |
| Your team has standardized on LangGraph or CrewAI and the existing setup works | ❌ | |
| You need fine-grained, code-first control over agent hierarchies and handoffs | ✅ | |
| You want a no-code or visual-builder experience for non-developers | ❌ |
Common Misconception
Myth: Google ADK only works with Gemini and locks you into Google Cloud. Reality: ADK is model-agnostic at the framework level — you can wire in non-Google models through standard interfaces — and the open-source library runs anywhere Python or TypeScript runs. Google Cloud’s Gemini Enterprise Agent Platform is an optional managed runtime, not a requirement.
One Sentence to Remember
If LangGraph is graphs and CrewAI is roles, ADK is Google’s code-first multi-agent toolkit with a paved road to managed deployment — pick it when the cloud destination is already decided, not before.
FAQ
Q: Is Google ADK the same as Vertex AI Agents? A: No. ADK is the open-source framework you install and run yourself. Vertex AI Agents was an earlier product; the current managed offering is the Gemini Enterprise Agent Platform, which runs ADK code.
Q: Does Google ADK support TypeScript? A: Yes. According to Google Developers Blog, a TypeScript SDK was added so frontend and full-stack teams can build agents in the same language as their application code, alongside the original Python SDK.
Q: How does Google ADK compare to LangGraph or CrewAI? A: All three are code-first multi-agent frameworks. ADK leans into Google Cloud and Gemini integration, LangGraph into explicit graph state machines, and CrewAI into role-based crews. Pick by ecosystem fit and orchestration style.
Sources
- Google Developers Blog: Agent Development Kit: Making it easy to build multi-agent applications - Original launch announcement and framework positioning.
- Google ADK Docs: Agent Development Kit (ADK) - Official documentation portal for the Python and TypeScript SDKs.
Expert Takes
ADK doesn’t change what an agent is — it’s still a model in a loop with tools and memory. What it changes is the engineering surface: typed primitives for agents, sub-agents, sessions, and evaluations, written in Python or TypeScript. The framework is a pattern library for the same orchestration problem LangGraph and CrewAI solve, expressed in Google’s vocabulary, with first-party hooks into Gemini’s tool-calling and reasoning behaviors.
Treat ADK like any other framework: the value is in the spec you write, not the SDK itself. Write a clear agent contract — instructions, tools, expected handoffs, evaluation criteria — and ADK becomes a faithful executor. Skip the spec and you’ll get the same drift you’d get from any other code-first toolkit. The local CLI and evaluation harness are there precisely so you can keep your spec honest before you deploy.
Google didn’t build ADK because the world was missing a multi-agent framework. They built it because every cloud now needs a flagship agent toolkit, and “use LangChain on our infra” isn’t a strategy. The pitch is paved-road simplicity: code locally, deploy to a managed runtime, stop gluing services together. If your team is already on Google Cloud, that path is real. If you’re vendor-neutral, the calculus is different.
A code-first framework owned by the cloud provider that also sells the runtime is a coordination story worth reading carefully. The open-source layer is genuine, and ADK runs anywhere. But the smooth path leads to one place. Ask yourself who decides when the framework’s defaults change, who sees the traces of your production agents, and how easy it actually is to walk away once your orchestration logic assumes a specific managed runtime.