LangGraph conditional edge to END does not stop execution

LangGraph conditional edge to END does not stop execution

TL;DR: if your graph keeps looping after the router says END, you almost certainly have BOTH an unconditional edge and a conditional edge leaving the same node. The unconditional edge always wins. Delete it and let the conditional edge decide. Also check you return exactly "END", not "END.".

# no exception. the symptom: should_continue returns "END" but the graph
# loops back through the generate node instead of stopping.

When this applies

• A conditional edge maps some return value to END.
• Execution continues past it anyway, usually looping one more node.
• The router function itself returns the right value (verify with a print).

When it does not

• GraphRecursionError after many loops means the router genuinely never returns END. Different problem.
• If the router raises, fix the router.

Fix it

1. Remove the competing unconditional edge

# wrong: this edge ALWAYS runs, so the conditional edge below never matters
graph.add_edge("eval", "generate")
graph.add_conditional_edges("eval", should_continue, {"CONTINUE": "generate", "END": END})

# right: only the conditional edge leaves "eval"
graph.add_conditional_edges("eval", should_continue, {"CONTINUE": "generate", "END": END})

Expected: when should_continue returns "END", the run stops. When it returns "CONTINUE", it goes to generate.

2. Return exactly "END", not "END."

def should_continue(state):
    if state["done"]:
        return "END"   # not "END." -- the period makes it a different node name
    return "CONTINUE"

Expected: the mapping key matches and routes to END. "END." matches nothing and (depending on version) either errors or falls through.

3. Verify with a print

def should_continue(state):
    decision = "END" if state["done"] else "CONTINUE"
    print("routing:", repr(decision))
    return decision

Expected: you see 'END' (with quotes, no period) on the final step, and the run stops right after.

Why it happens

Edges out of a node ALL fire unless they are conditional. An unconditional add_edge("eval", "generate") plus a conditional edge from "eval" means generate runs every time AND the conditional branch runs: the graph does both. Newcomers add the unconditional edge first (from a tutorial) and then add the conditional edge without removing it.

Edge cases

• With interrupt_before/interrupt_after on the same node, the interrupt fires but the extra edge still routes afterward. Clean the edges first.
• Returning END (the constant) vs "END" (the string): both work in add_conditional_edges mappings, but be consistent. Mixing them across the router and the mapping is a classic miss.
• If two conditional edges leave the same node, both are evaluated. That is legal but confusing; prefer one router.

Compatibility

All langgraph Python versions, 0.x and 1.x.