DEV Community

Jangwook Kim
Jangwook Kim

Posted on • Originally published at jangwook.net

Google ADK vs LangGraph 2026: I Installed Both and Compared Them Side by Side

Every time a new AI agent framework drops, my first instinct is to install it and figure out what's actually different. When Google open-sourced ADK (Agent Development Kit), I set up a clean sandbox environment and ran Google ADK v1.32.0 alongside LangGraph v1.1.10. This post is what I found.

Sandbox execution logs comparing ADK and LangGraph

Two Different Philosophies

Google ADK's core message is "apply software development principles to AI agents." That becomes concrete the moment you write code. ADK defines agents as Python classes and functions. Built-in orchestrators — SequentialAgent, ParallelAgent, LoopAgent — let you declare flow naturally inside your code.

from google.adk.agents import Agent, SequentialAgent, ParallelAgent

weather_agent = Agent(
    name="weather_agent",
    model="gemini-2.5-flash",
    instruction="Get weather data using the get_weather tool.",
    tools=[get_weather],
)

analysis_agent = Agent(
    name="analysis_agent",
    model="gemini-2.5-flash",
    instruction="Analyze the weather data and give a forecast.",
)

# Sequential execution: weather → analysis
pipeline = SequentialAgent(
    name="weather_pipeline",
    sub_agents=[weather_agent, analysis_agent],
)
Enter fullscreen mode Exit fullscreen mode

No separate graph definition, no edge declarations. The Python code itself expresses the flow.

LangGraph takes the opposite approach. It models agent workflows as explicit graphs. Nodes represent processing stages; edges define transitions between them. You design the graph first, then add logic on top.

from typing import TypedDict, Annotated
from langgraph.graph import StateGraph, END
from langgraph.graph.message import add_messages
from langchain_core.messages import HumanMessage, AIMessage

class State(TypedDict):
    messages: Annotated[list, add_messages]

def greet_node(state: State):
    return {"messages": [AIMessage(content="Hello from LangGraph!")]}

def analyze_node(state: State):
    last = state["messages"][-1].content
    return {"messages": [AIMessage(content=f"Analysis: {last}")]}

builder = StateGraph(State)
builder.add_node("greet", greet_node)
builder.add_node("analyze", analyze_node)
builder.set_entry_point("greet")
builder.add_edge("greet", "analyze")
builder.add_edge("analyze", END)

graph = builder.compile()
Enter fullscreen mode Exit fullscreen mode

The way I see it: ADK says "assemble a pipeline in code"; LangGraph says "design a state machine." Neither is objectively better — it's about which feels more natural for your problem.

The Dependency Gap Is Shocking 

pip install google-adk langgraph
Enter fullscreen mode Exit fullscreen mode

Both install fine, but pip show reveals a stark difference:

Google ADK v1.32.0 — 45 direct dependencies:

aiosqlite, anyio, authlib, click, fastapi,
google-api-python-client, google-auth,
google-cloud-aiplatform, google-cloud-bigquery,
google-cloud-bigquery-storage, google-cloud-bigtable,
google-cloud-dataplex, google-cloud-discoveryengine,
google-cloud-pubsub, google-cloud-secret-manager,
google-cloud-spanner, google-cloud-speech,
google-cloud-storage, google-genai, graphviz,
httpx, jsonschema, mcp, opentelemetry-api,
opentelemetry-exporter-gcp-logging,
opentelemetry-exporter-gcp-monitoring,
opentelemetry-exporter-gcp-trace,
... (45 total)
Enter fullscreen mode Exit fullscreen mode

LangGraph v1.1.10 — 6 direct dependencies:

langchain-core, langgraph-checkpoint,
langgraph-prebuilt, langgraph-sdk,
pydantic, xxhash
Enter fullscreen mode Exit fullscreen mode

A 39-package gap. ADK bundles Google Cloud services (BigQuery, Spanner, Pub/Sub, Speech, etc.), OpenTelemetry exporters, a FastAPI server, and SQLAlchemy ORM. If you're not using Google Cloud, those 39 extra packages are pure dead weight.

LangGraph's "bring what you need" philosophy is lighter but requires more configuration. You inject your own LLM client, choose your own checkpoint backend. Less magic, more control.

Multi-Agent Patterns: Where the Difference Matters Most

ADK parallel execution — this is actual code I ran:

parallel_research = ParallelAgent(
    name="parallel_research",
    sub_agents=[google_researcher, arxiv_researcher],
)

pipeline = SequentialAgent(
    name="research_pipeline",
    sub_agents=[parallel_research, synthesizer],
)
Enter fullscreen mode Exit fullscreen mode

Execution: research_pipeline (SequentialAgent) → parallel_research (ParallelAgent) → synthesizer. Clean, readable, obvious.

LangGraph conditional edges — this is where ADK falls short:

def should_retry(state: State) -> str:
    if state.get("quality_score", 0) < 80:
        return "generate"  # quality too low → retry
    return END              # quality passed → done

builder.add_conditional_edges("evaluate", should_retry)
Enter fullscreen mode Exit fullscreen mode

I ran this in my sandbox:

=== Conditional Edge Test ===
  [evaluate] iteration=1, score=60  → routed to: generate
  [evaluate] iteration=2, score=80  → routed to: END
Final score: 80, total iterations: 2
Enter fullscreen mode Exit fullscreen mode

The agent retries automatically until quality passes. ADK's LoopAgent supports iteration too, but termination is controlled by max_iterations. Dynamic branching — "if this condition, go here; otherwise go there" — is where LangGraph's conditional edges shine. For production pipelines with generate-verify-retry loops, router agents, or multi-judgment branching, LangGraph is significantly more powerful.

ADK's Killer Feature: The CLI and Built-in Evaluation

Installing ADK drops the adk CLI:

$ adk --help
Commands:
  api_server   Starts a FastAPI server for agents.
  conformance  Conformance testing tools for ADK.
  create       Creates a new app with prepopulated code.
  deploy       Deploys agent to hosted environments.
  eval         Evaluates an agent given eval sets.
  eval_set     Manage Eval Sets.
  migrate      ADK migration commands.
  optimize     Optimizes root agent instructions (GEPA).
  run          Runs an interactive CLI for an agent.
  web          Starts a FastAPI server with Web UI.
Enter fullscreen mode Exit fullscreen mode

adk web is particularly useful. Point it at your agent code and a FastAPI-backed web UI spins up for local visual testing. adk eval lets you define evaluation sets and run regression tests against your agent automatically. LangGraph has none of this built in.

adk deploy supports direct deployment to Google Cloud Run or Vertex AI Agent Builder. If you're on GCP, development to deployment stays within a single tool chain.

The downside: this CLI is entirely Google-ecosystem bound. adk deploy is useless on AWS or Azure. The built-in tracing outputs to GCP Cloud Trace — connecting to other observability stacks requires extra configuration.

That's where a vendor-agnostic tool like Langfuse for LLM tracing fits more naturally with LangGraph's approach.

State Management: Sessions vs Checkpoints

ADK's state management is session-based:

from google.adk.runners import InMemoryRunner
from google.adk.sessions import InMemorySessionService

session_service = InMemorySessionService()
runner = InMemoryRunner(agent=root_agent, app_name="my_app")
Enter fullscreen mode Exit fullscreen mode

ADK maintains multi-turn conversation state via session_id. In production, swap InMemorySessionService for VertexAiSessionService for persistence. Inter-agent data flows through a shared State dictionary.

LangGraph's state management uses TypedDict + Checkpoint:

from langgraph.checkpoint.memory import MemorySaver

checkpointer = MemorySaver()
graph = builder.compile(checkpointer=checkpointer)

config = {"configurable": {"thread_id": "user-123"}}
result = graph.invoke({"messages": [HumanMessage(content="Hello")]}, config)
Enter fullscreen mode Exit fullscreen mode

LangGraph's checkpoint system is more flexible. Swap MemorySaver for PostgreSQL, Redis, or SQLite backends. Time-travel debugging — rewinding to a past checkpoint — is also supported.

I consider LangGraph stronger here. Swapping checkpoint backends means dev and production can run identical code. No vendor lock-in.

ADK does beat LangGraph on one point: MCP tool server support is built in. MCPToolset works out of the box. With LangGraph, you need a separate package and adapter code.

Comparison Table

Aspect Google ADK v1.32.0 LangGraph v1.1.10
Direct dependencies 45 6
Orchestration style Sequential/Parallel/LoopAgent (declarative) StateGraph + nodes/edges (explicit graph)
Conditional branching LoopAgent max_iterations (limited) conditional_edges (powerful)
Default LLM Gemini (others supported) Model-agnostic (inject any LLM)
CLI adk create/run/web/eval/deploy ✓ None
Built-in Web UI adk web ✓ None
Built-in eval framework adk eval ✓ None (external tooling needed)
MCP support MCPToolset built-in ✓ Separate package needed
State management Session-based (VertexAI backend) TypedDict + Checkpoint (swappable backends)
Deployment Google Cloud Run / Vertex AI Cloud-agnostic
OpenTelemetry GCP exporters built-in Manual setup
Time-travel debugging No ✓
Multi-language SDK Python, Go, Java, TypeScript Python (primary)
License Apache 2.0 MIT

Which Framework for Which Team

Google ADK fits when:

  • Your infrastructure is already on Google Cloud
  • Gemini is your primary model
  • You want development-to-deployment in a single toolchain
  • You don't want to build an agent evaluation pipeline from scratch
  • Your team works in Go, Java, or TypeScript alongside Python

LangGraph fits when:

  • Your agent workflows have complex branching logic
  • You're on AWS, Azure, or multi-cloud
  • You mix multiple LLMs (OpenAI, Anthropic, Mistral, etc.)
  • You have existing LangChain-based code
  • You need time-travel debugging or checkpoint replay
  • You're minimizing dependencies for edge deployments

If I were starting a new project today, I'd reach for LangGraph. The reason is pragmatic: conditional branching and checkpoint flexibility become necessary in production-grade agents, and retrofitting them later often means redesigning the whole structure. Starting with ADK's LoopAgent and needing dynamic routing later could force a significant rewrite.

That said, for teams already on GCP, the ADK toolchain is genuinely compelling. adk deploy to Cloud Run plus adk eval for regression testing would take real work to replicate in the LangGraph ecosystem.

My Take: Design Philosophy Decides the Choice

Both frameworks are production-ready in 2026. The difference is what they optimize for.

ADK optimizes for "build an agent system fast and deploy it to GCP." LangGraph optimizes for "precisely control the state transitions in your agent."

The mistake I see in most framework comparisons is ranking by feature count. ADK has a CLI, eval, and deployment — so it wins? No. The real question is: does the framework's growth direction match my agent's complexity growth direction?

If a simple pipeline will evolve toward complex branching, choose LangGraph early. If the goal is fast delivery within the Google Cloud ecosystem, ADK removes a lot of friction.

For broader context on how LangGraph compares with other frameworks, the LangGraph vs CrewAI vs Dapr comparison is worth reading — it benchmarks them on production KPIs before ADK entered the picture.

Top comments (0)