Skip to content
Short answer: In the Google ADK vs LangGraph choice, pick ADK if you want an agent framework in Python, TypeScript, Go, Java or Kotlin with an easy path to Google Cloud. Pick LangGraph if you want fine control over long-running, stateful agents written as graphs in Python or JavaScript. Both are free and open source: ADK is Apache 2.0 and LangGraph is MIT.
Both frameworks help you build agents that call tools, keep state and work together. Our Google ADK page and LangGraph page cover each one on its own. This post puts them side by side so you can see where they differ and which one suits your team.
Google ADK vs LangGraph at a glance
| Google ADK | LangGraph | |
|---|---|---|
| What it is | Open-source, code-first framework for building, debugging and deploying agents | Low-level orchestration framework and runtime for stateful agents |
| Made by | LangChain | |
| License | Apache 2.0 | MIT |
| Languages | Python, TypeScript, Go, Java, Kotlin | Python and JavaScript/TypeScript |
| Install (Python) | pip install google-adk | pip install -U langgraph |
| Models | Almost any model, optimized for Gemini | Works with any model through its integrations |
| Workflows | Sequential, loop and parallel agents, plus graph workflows in ADK 2.0 | Graphs of nodes with persistent state |
| Human in the loop | Tool confirmation and workflow support | Inspect and change agent state at any point |
| Cost | Free framework, model usage extra | Free framework, model usage extra |
What is Google ADK?
Google describes the Agent Development Kit as an open-source framework that lets you build, debug and deploy reliable AI agents. It is code-first, so you define agent logic and orchestration in your own code. A no-code Agent Config option also exists for simple agents.
Language support is its broadest feature. The ADK docs list Python, TypeScript, Go, Java and Kotlin, each with its own install command. If your team writes backend services in Go or Java, you can build agents without switching languages.
For multi-agent work, ADK gives you agent routing, sub-agent delegation and template workflows for sequential, loop and parallel agents. ADK 2.0 adds graph workflows that mix deterministic code with AI reasoning along explicit paths. It also supports tools from built-in libraries, custom functions, OpenAPI specs and MCP servers. If MCP is new to you, our guide to what MCP servers are explains it.
What is LangGraph?
LangGraph is a low-level orchestration framework and runtime for building stateful agents. You model an agent as a graph: nodes do the work and edges decide what runs next. That lets you combine hand-written logic with model-driven decisions in one flow.
Its main selling points are the ones long-running agents need. Durable execution lets an agent survive a failure and resume where it stopped. You can inspect and change an agent’s state while it runs. It supports short-term working memory and long-term memory across sessions.
LangGraph does not require LangChain. LangChain’s docs say you do not need to use LangChain to use LangGraph, though the two integrate. If you are weighing the two, read LangChain vs LangGraph.
Language and model support
This is the clearest split. ADK covers five languages. LangGraph covers two: Python, plus JavaScript and TypeScript through LangGraph.js. If you are a Python or TypeScript shop, both fit and the language will not decide it. If you work in Go, Java or Kotlin, ADK is the only one of the two that speaks your language.
On models, both are open. ADK says it works with almost any generative AI model, with easy access to Gemini, adapters for other providers and support for locally running models. Its README calls it model-agnostic and optimized for Gemini. LangGraph leaves model choice to you and works with model providers through its integrations. Neither locks you to one vendor, but ADK is tuned for Google’s own models.
Control and workflow design
LangGraph is deliberately low level. You decide the state, the nodes and the edges. The framework gives you persistence, memory and human checkpoints. That is a lot of control. The cost is a steeper learning curve, which is the main complaint we note on our LangGraph page.
ADK offers more ready-made patterns first. You can start with a sequential, loop or parallel agent. Or you can let a parent agent delegate to sub-agents. Then move to graph workflows when you need explicit paths. Its Workflow Runtime handles routing, loops, retries and state. Its Tool Confirmation flow asks for explicit approval before a tool runs.
Put simply, LangGraph starts from the graph and ADK starts from agent patterns and adds graphs later. If you already know how you want the flow to look, LangGraph’s model is direct. If you want a working multi-agent setup quickly, ADK’s patterns get you there faster.
Deployment and cost
Both frameworks are free to use, and both leave you paying for the models you call.
ADK deployment
ADK runs in containers on your own infrastructure. It also deploys to Google Cloud through Agent Runtime, Cloud Run or GKE. Managed hosting on Google Cloud is optional and billed by usage under Google Cloud’s pricing. Most deployment guides focus on Google Cloud, so expect the smoothest path there.
LangGraph deployment and LangSmith
LangGraph itself costs nothing. The paid part is LangSmith, LangChain’s platform for tracing, evaluation and hosting.
| LangSmith plan | Price | What it includes |
|---|---|---|
| Developer | $0 per seat a month | 1 seat, 5k base traces a month, then pay as you go |
| Plus | $39 per seat a month | 10k base traces a month, one free small serverless deployment |
| Enterprise | Custom | Self-hosted and hybrid options |
Hosted deployment of LangGraph agents needs Plus or higher, and deployment billing is usage based, so it is harder to predict. Seats are billed monthly and traces are billed monthly in arrears. You can still run LangGraph on your own servers without LangSmith.
A first project in each
With ADK you install the package for your language, define an agent with a model and a few tools, and run it from the CLI or the local dev UI. To add a second agent, you attach it as a sub-agent or place both in a sequential or parallel workflow. You grow the design in steps and add graph workflows only if the flow needs them.
With LangGraph you install the package, define the state your agent carries, add nodes and connect them with edges. It takes more setup before the first run. In return you know exactly what happens at each step, where state is saved and where a person can step in. That clarity matters when an agent runs for hours and must not lose its place.
A fair test is to build the same small agent in both, such as one that researches a topic and drafts a summary. Compare how long it takes to get a first result, how easy the flow is to read afterward and how each handles a failed tool call.
Where Google ADK wins
- Language range. Python, TypeScript, Go, Java and Kotlin.
- Ready-made patterns. Sequential, loop and parallel agents and sub-agent delegation work out of the box.
- Google Cloud path. Agent Runtime, Cloud Run and GKE are documented deployment targets.
- No-code option. Agent Config builds simple agents without writing code.
Where LangGraph wins
- Fine control. You shape the state, nodes and edges yourself.
- Durable execution. Agents resume after failures and run for long periods.
- State you can inspect. You can look at and change agent state mid-run.
- Tracing with LangSmith. Debugging and evaluation tools sit next to the framework.
What neither does
Neither framework writes a good agent for you. You still design the tools, the prompts and the checks. You also need to test what the agent does. Neither removes model costs, and neither guarantees correct answers. Both also change quickly, so pin your versions and read release notes before upgrading. For a wider view, see our best AI agent frameworks page, or read how multi-agent systems work.
Which should you choose?
- You write Go, Java or Kotlin: ADK.
- You deploy on Google Cloud: ADK.
- You want ready-made multi-agent patterns: ADK.
- You need long-running agents that resume after failures: LangGraph.
- You want the lowest-level control over state and flow: LangGraph.
- You already use LangChain tools or LangSmith: LangGraph.
- You are new to agents: start with ADK’s patterns, or read how to build an AI agent first.
FAQ
Is Google ADK better than LangGraph?
Neither is better for everyone. ADK is stronger on language support and ready-made multi-agent patterns. LangGraph is stronger on low-level control and durable, stateful execution.
Is Google ADK free?
Yes. The framework is open source under Apache 2.0. You pay for the models you call, and for Google Cloud hosting if you choose managed deployment.
Is LangGraph free?
Yes. LangGraph is open source under the MIT license. LangSmith’s Developer plan is $0 per seat a month, and Plus is $39 per seat a month for hosted deployment.
Can Google ADK use models other than Gemini?
Yes. ADK says it works with almost any generative AI model and offers adapters for other providers and support for local models. It is optimized for Gemini.
Does LangGraph need LangChain?
No. LangChain’s docs say you do not need LangChain to use LangGraph, though it integrates with LangChain products.
Which supports more programming languages?
Google ADK. It supports Python, TypeScript, Go, Java and Kotlin, while LangGraph supports Python and JavaScript/TypeScript.
Keep reading
Details here come from the Google ADK docs, the LangGraph docs and LangChain’s pricing page, checked in October 2026. To compare more frameworks, browse the Klyvoo AI agent frameworks page or the MCP servers directory.