Migration from LangGraph

Migration

From LangGraph boilerplate to Nodex

Keep the graph model, remove repetitive wiring, and get a cleaner execution trace automatically. The two code blocks below show the same two-node graph written in raw LangGraph and in Nodex.

LangGraph (before)

from langgraph.graph import END, START, StateGraph

def research(state):
    return {"notes": state["query"]}

def writer(state):
    return {"final": state["notes"].upper()}

graph = StateGraph(dict)
graph.add_node("research", research)
graph.add_node("writer", writer)
graph.add_edge(START, "research")
graph.add_edge("research", "writer")
graph.add_edge("writer", END)
app = graph.compile()
app.invoke({"query": "LangGraph"})

Nodex (after)

from nodex import Agent

app = Agent()

@app.node(next="writer")
def research(state):
    return {"notes": state.get("query")}

@app.node(next="end")
def writer(state):
    return {"final": state.get("notes").upper()}

app.run({"query": "LangGraph"})

Removed ceremony

What you no longer need to write

Nodex handles this wiring internally. You can delete it from your existing agent code.

StateGraph(dict)	Nodex creates the underlying LangGraph graph automatically when you call `app.run()`.
graph.add_node(name, fn)	Replaced by `@app.node()`. The function name becomes the node name.
graph.add_edge(A, B)	Replaced by `next="node_name"` on the decorator.
graph.add_edge(START, first)	Nodex infers the first node automatically from registration order.
graph.add_edge(last, END)	Replaced by `next="end"` on the last node.
graph.compile()	Called automatically inside `app.run()`.
Custom retry loops	Replaced by `retry=N` and `on_fail="node_name"` on `@app.node()`.

Checklist

Migration checklist

Nodes	Move each LangGraph node function under `@app.node`.
Edges	Replace direct edges with `next="node_name"`.
END	Use `next="end"` for terminal nodes.
Conditional edges	Use `@app.route(condition, if_true, if_false)`.
State reads	Replace `state["key"]` with `state.get("key")`.
Retries	Move retry logic from try/except wrappers to `retry=N, on_fail=` on the decorator.
Tests	Test one full graph path and key nodes in isolation with `test_node()`.

Escape hatch

Still need raw LangGraph?

Nodex is a layer, not a lock-in. If you need advanced LangGraph features — subgraphs, streaming, checkpointers — you can still use them directly. Nodex compiles down to a standard LangGraph graph, so you can drop back to the LangGraph API for anything not covered by Nodex decorators.

The LangGraph documentation covers advanced features like streaming, persistence, and human-in-the-loop at the framework level.