LangGraph vs Google ADK: which agent framework to use?

Pick LangGraph if you want to design the control flow yourself and need runs that survive failures, pause for a human and resume from a checkpoint. Pick Google ADK if you want an agent running quickly from a model, an instruction and a list of tools, need Go, Java or Kotlin, or plan to deploy on Google Cloud. Both are open source and both now do graph workflows, so the real split is how low-level you want to go. Sources checked 2026-10-27.

Quick comparison

How they differ

Level of abstraction

LangGraph's docs call it "very low-level" and "focused entirely on agent orchestration". It does not abstract prompts or agent architecture, and the docs recommend LangChain's prebuilt agents to anyone who wants something higher-level. In practice you write a state schema, nodes and edges, then compile the graph.

ADK goes the other way. Its site's first example is a complete research agent in a few lines: a name, a model, an instruction and the google_search tool. You add structure later. The site describes the path as prompts and tool calls first, then multi-agent orchestration, graph workflows, evaluation and deployment.

Workflows

In LangGraph everything is a graph. Its main selling point in the docs is mixing deterministic, hand-coded steps with LLM-driven steps in the same graph, so some parts stay predictable and auditable.

ADK lists four workflow types. Graph-based, dynamic (plain code logic) and collaborative (one coordinator agent with sub-agents) workflows need ADK 2.0 or later; template workflows give fixed sequence, loop and parallel patterns. Agent routing between agents at runtime is marked experimental.

State, persistence and human review

This is LangGraph's strongest area. The docs list persistence so agents "persist through failures and can run for extended periods, resuming from where they left off", human-in-the-loop inspection and editing of state at any point, and short- and long-term memory. Recent releases keep tightening it: 1.2.18 streams the checkpoint a run stopped on, so resumed clients no longer repeat the last node.

ADK focuses on context rather than durability. Its FAQ says sessions, memory, tool outputs and artifacts are assembled into a structured view, with old turns summarized and token use tracked. Human approval exists as a workflow feature: adk-python 2.12.0 added require_confirmation for nodes.

Models

LangGraph has no model layer of its own. The docs use LangChain integrations for models and tools, though LangChain is optional.

ADK's FAQ says it works with almost any generative model, giving easy access to Gemini and adapters for other providers, including local models. All of the site's examples use gemini-flash-latest.

Deployment and tooling

LangGraph pairs with LangSmith for tracing, evaluation and a managed deployment platform for long-running, stateful workflows. The cited sources do not give LangSmith pricing.

ADK can be containerized and run anywhere, or deployed with one command to Agent Runtime (Agent Platform), Cloud Run or GKE. The site says agents then get managed infrastructure, built-in authentication and Cloud Trace observability without code changes. It also has an Agents CLI to scaffold, test, evaluate and deploy.

Which one to choose

• Long-running, resumable agents with human checkpoints: LangGraph.
• Teams in Go, Java or Kotlin: ADK; LangGraph has no official library for them.
• Already on Google Cloud and Gemini: ADK, for the one-command deployment path.
• You want a working agent before designing a graph: ADK, or LangChain's agents on top of LangGraph.
• Strict control over every step and state transition: LangGraph.

Who should skip each

Skip LangGraph if you want batteries included: you will be assembling the agent loop yourself or adding LangChain. Skip ADK if your deployment target is not Google Cloud and durable resumable execution is a hard requirement, since the cited ADK sources do not describe it.

How we compared

We read the LangGraph overview in the LangChain docs, the adk.dev home page and FAQ, the ADK workflows page, GitHub repository metadata for langgraph, langgraphjs and the five ADK repositories, PyPI metadata and the latest release notes for both projects, all on 2026-10-27. We did not run benchmarks or deploy either framework.

Limitations

Both descriptions come from the vendors' own sites. Neither source gives pricing for LangSmith or Google Cloud hosting, and we did not compare feature parity across ADK's five languages or between LangGraph's Python and JavaScript libraries.

Related reading

Recent launches: LangGraph 1.2.18 and adk-python 2.13.0. Other comparisons: Agno vs LangGraph, Pydantic AI vs LangGraph, DSPy vs LangGraph, Pydantic AI vs Google ADK and Agno vs Google ADK. Browse the frameworks category.