LangGraph Fundamentals

LangGraph is LangChain's framework for complex, stateful agent workflows. While LCEL suffices for linear chains, LangGraph enables cycles, conditional branching, and persistent state — essential for production agents.

Why LangGraph?

Simple chains are linear: Input → Processing → Output. But real agent workflows need:

• Cycles: The agent tries something, checks the result, and tries again
• Conditional logic: Different paths are chosen based on results
• Persistent state: State is maintained across multiple steps
• Human-in-the-loop: Humans can intervene at defined points

State Graph Concept

A LangGraph graph consists of three elements:

1. State

The shared state that flows through the graph:

from typing import TypedDict, Annotated
from langgraph.graph.message import add_messages

class AgentState(TypedDict):
    messages: Annotated[list, add_messages]
    current_step: str
    results: dict

2. Nodes

Functions that process and modify state:

def research_node(state: AgentState) -> dict:
    # Perform research
    results = search_tool.invoke(state["messages"][-1].content)
    return {"results": {"research": results}, "current_step": "analyze"}

3. Edges

Connections between nodes — static or conditional:

from langgraph.graph import StateGraph, END

graph = StateGraph(AgentState)
graph.add_node("research", research_node)
graph.add_node("analyze", analyze_node)
graph.add_node("respond", respond_node)

graph.add_edge("research", "analyze")
graph.add_conditional_edges("analyze", route_function, {
    "needs_more": "research",   # Cycle back
    "ready": "respond"          # Continue to respond
})
graph.add_edge("respond", END)

Conditional Routing

Conditional edges allow dynamic decisions:

def route_function(state: AgentState) -> str:
    if state["results"].get("confidence", 0) < 0.8:
        return "needs_more"
    return "ready"

Checkpointing

LangGraph saves state after every node. On errors, you can roll back to the last checkpoint:

from langgraph.checkpoint.memory import MemorySaver

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

Human-in-the-Loop

Define interrupt points where human approval is required:

app = graph.compile(
    checkpointer=checkpointer,
    interrupt_before=["send_email", "delete_record"]
)

Practical tip: Draw your workflow as a flowchart before writing code. Every box becomes a node, every arrow an edge. LangGraph makes your architecture explicit — use that.