LangGraph vs LangChain

How They Relate

LangChain is a framework for building LLM apps (prompts, chains, tools, memory, agents). LangGraph is a library built on top of LangChain that adds graph-based control flow so you can build complex, stateful, loop‑capable workflows.

State Graph: The Mental Model

A state graph is like a state machine specialized for AI systems. A structured state object flows through the graph while nodes do work and edges decide what runs next. Nodes do not live in the state; they read from it and write to it.

[UserInput] → [Analyze] → [Decide]
                          │
              if search_needed ─────▶ [WebSearch] ──▶ [Synthesize]
                          └──────────▶ [AnswerDirectly]

Nodes

A node is a unit of computation: a LangChain chain/agent, a plain function, or even a nested subgraph. Each node takes the current state, performs work, and returns an updated state.

# pseudo-Python
def search_node(state):
    query = state["analysis"]["query"]
    results = web_search(query)
    state["search_results"] = results
    return state

Edges

Edges implement control flow: if/then/else, branching, and loops. After a node updates state, an edge inspects it to choose the next node.

Checkpointing and Persistence

Checkpointing saves a snapshot of progress and state right after a node finishes. That enables pause/resume, replay, and fault recovery. Persistence is the backend that stores those snapshots (memory, SQLite, Postgres, Redis, etc.).

# conceptual API
from langgraph.checkpoint.memory import MemorySaver
# or:
from langgraph.checkpoint.sqlite import SqliteSaver

checkpointer = SqliteSaver("graph_state.sqlite")
graph = build_graph(nodes, edges, checkpointer=checkpointer)

thread_id = "customer-123"
graph.run(input_state, thread_id=thread_id)  # checkpoints after each node

Threading

Thread = one storyline of state. Each conversation or task gets its own thread with its own checkpoint history. The same graph definition can run many threads concurrently—exactly what you want for production agents.

Example: Support Agent Graph

1. ClassifyIntent
2. SearchKnowledgeBase
3. GenerateAnswer
4. EscalateToHuman

Edges route “simple” intents to search/answer, and “complex” intents to escalation. After every node, the graph checkpoints; if the process restarts, it resumes from the last completed node, not the beginning.