LangGraph Production Deployment

LangGraph from LangChain treats agent workflows as graphs of states and transitions rather than free-form conversation loops. Explicit, debuggable, more predictable than ReAct-style agents. On our dedicated GPU hosting it pairs well with self-hosted LLMs for production deployments.

Contents

• Why LangGraph
• LLM config
• Graph example
• Persistence and checkpoints

Why LangGraph

Free-form agents (ReAct) are powerful but unpredictable. LangGraph forces you to model the workflow explicitly: “from state A go to state B if condition X.” This makes failures localisable and debugging tractable. Good fit for structured production workflows; heavy-weight for simple chatbots.

LLM Config

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="llama-3.3-70b",
    openai_api_base="http://localhost:8000/v1",
    openai_api_key="not-needed",
    temperature=0.1,
)

Graph

from langgraph.graph import StateGraph, END

def classify(state):
    # run LLM to classify the intent
    return {"intent": classification}

def answer_factual(state):
    # RAG pipeline
    return {"answer": answer}

def answer_action(state):
    # tool-use pipeline
    return {"answer": answer}

graph = StateGraph(dict)
graph.add_node("classify", classify)
graph.add_node("factual", answer_factual)
graph.add_node("action", answer_action)
graph.set_entry_point("classify")
graph.add_conditional_edges("classify", lambda s: s["intent"])
graph.add_edge("factual", END)
graph.add_edge("action", END)

app = graph.compile()
result = app.invoke({"query": user_query})

Persistence

LangGraph supports checkpointed state – pause a workflow, resume it, replay from any step. Useful for long-running agent tasks where the user may return hours later. Checkpoint to PostgreSQL or SQLite depending on scale.

See AutoGen and CrewAI.