Five options

Raw SDK · LangChain · LangGraph · CrewAI · OpenClaw

LangChain vs LangGraph

RAG chain

retriever | prompt | llm | parser

RAG code

rag_chain = (
  {"context": retriever, "question": RunnablePassthrough()}
  | prompt | llm | StrOutputParser()
)
rag_chain.invoke("Refund window?")

ReAct graph

StateGraph · ToolNode · conditional_edges

Graph code

g = StateGraph(AgentState)
g.add_node("agent", call_model)
g.add_node("tools", ToolNode([search_docs]))
g.add_conditional_edges("agent", tools_condition)
app = g.compile(checkpointer=SqliteSaver(...))

Human-in-loop

interrupt_before=["tools"]

SqliteSaver

cfg = {"configurable": {"thread_id": "u1"}}
app.invoke({"messages": [...]}, cfg)
# same thread_id later → remembers ORD-99

Intent router

conditional_edges(START, route_intent)

order → order_agent · else → rag_agent

Checkpoint vs Store

Support agent lab

1. LangChain RAG baseline
2. LangGraph + tools
3. SqliteSaver same thread
4. Store cross-thread
5. interrupt_before refund

Scroll: #lab + full code