Langgraph: Conditional 'END' State Fails to Stop Execution

[fzanart]

I am using Langgraph to create a state graph, but I'm encountering an issue where the END condition in my conditional edge is not terminating the execution as expected. Instead of stopping at the END node, the graph continues to loop through the generate node even when should_continue returns "END."

Pyhton 3.12.4
Langgraph 0.2.12

Here is my reproducible code:

import operator
import logging
import random
from typing import TypedDict, Annotated, Sequence, List
from langgraph.graph import StateGraph, END
from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage

logging.basicConfig(level=logging.INFO)


class AgentState(TypedDict):
    messages: Annotated[Sequence[BaseMessage], operator.add]


class Agent:
    def __init__(self):
        # graph init
        graph = StateGraph(AgentState)
        graph.add_node("init", self.init_node)
        graph.add_node("generate", self.generate_node)
        graph.add_node("eval", self.eval_node)
        graph.add_node("search", self.search_node)

        graph.set_entry_point("init")
        graph.add_conditional_edges(
            "eval",
            self.should_continue,
            {
                "END": END,
                "CONTINUE": "generate",
            },
        )
        graph.add_edge("init", "search")
        graph.add_edge("search", "eval")
        graph.add_edge("eval", "generate")
        graph.add_edge("generate", "search")

        self.graph = graph.compile()

    def init_node(self, state: AgentState):

        content = "some keywords"
        logging.info(content)
        return {"messages": [AIMessage(content=content)]}

    def search_node(self, state: AgentState):

        content = "some results...."
        logging.info(content)
        return {"messages": [AIMessage(content=content)]}

    def eval_node(self, state: AgentState):

        content = "some evaluation...."
        logging.info(content)
        return {"messages": [AIMessage(content=content)]}

    def generate_node(self, state: AgentState):

        content = "some new keywords...."
        logging.info(content)
        return {"messages": [AIMessage(content=content)]}

    def should_continue(self, state: AgentState):

        choice = random.choice(["END", "CONTINUE"])
        logging.info(choice)
        return choice

Then I call it like this:

from reproducible_agent import Agent
from langchain_core.messages import HumanMessage

bot = Agent()
messages = [HumanMessage(content="hello world")]
result = bot.graph.invoke({"messages": messages})

logs:

INFO:root:some keywords. INFO:root:some results....
INFO:root:some evaluation....
INFO:root:END. INFO:root:some new keywords....
INFO:root:some results....
INFO:root:some evaluation....
INFO:root:END. INFO:root:some new keywords....
INFO:root:some results....
INFO:root:some evaluation....
INFO:root:CONTINUE. INFO:root:some new keywords....
INFO:root:some results....
INFO:root:some evaluation....
INFO:root:CONTINUE. INFO:root:some new keywords....
INFO:root:some results....
INFO:root:some evaluation....
INFO:root:CONTINUE. INFO:root:some new keywords....
INFO:root:some results....
INFO:root:some evaluation....
INFO:root:CONTINUE. INFO:root:some new keywords....
...
INFO:root:some results....
INFO:root:some evaluation....
INFO:root:CONTINUE. INFO:root:some new keywords....

Eventually, it breaks with the following error:

GraphRecursionError: Recursion limit of 25 reached without hitting a stop condition. You can increase the limit by setting the recursion_limit config key.

I also tried, with no luck

...
graph.add_conditional_edges(
            "eval",
            self.should_continue)

...

    def should_continue(self, state: AgentState):

        choice = random.choice(["END", "CONTINUE"])
        logging.info(choice)

        if choice == "END":
            return END
        else:
            return "generate"

[jacoblee93]

Confusingly, END != "END" - can you try using the full value of __end__ instead where you do string comparisons?

[fzanart]

It doesn't work either

[fzanart]

I removed graph.add_edge("eval", "generate") and it worked. Thanks

[hinthornw]

Yupp! Edges are all run / are additive. Returning __end__ doesn't really take an action (it doesn't cancel in flight tasks) it just means "don't schedule a new task" and is equivalent in general to returning nothing / an empty list from a conditional list.

Since you previously were also always transitioning from eval to generate, this is like instructing the scheduler: "Next nodes are None and generate".

In sum, it's more correct to think of state graph as transitioning to all nodes that are scheduled by outgoing edges. It will not let one edge override the others by sending "end" or anything else.