Getting Multiple Interrupts at the same time from graph

OSS Product Help LangGraph
1

While developing an application using LangGraph, I wanted to implement a human-in-the-loop feature using LangGraph’s interrupt mechanism.
My setup includes a main graph and a subgraph, and both contain one interrupt for human feedback.
Below is the code I’m using to compile the graphs:

from langgraph.graph import StateGraph, START, END

# --- Part 1: Build the Response Agent Subgraph ---

# Initialize a new state graph with our defined `State` schema.
agent_builder = StateGraph(State)

# Add nodes for LLM call and interrupt handling.
agent_builder.add_node("llm_call", llm_call)
agent_builder.add_node("interrupt_handler", interrupt_handler)

# Define graph flow: start → llm_call
agent_builder.add_edge(START, "llm_call")

# Add conditional branching: 
# If continuation is needed, go to interrupt_handler; otherwise, end.
agent_builder.add_conditional_edges(
    "llm_call",
    should_continue,
    {
        "interrupt_handler": "interrupt_handler",
        END: END,
    },
)

# After handling an interrupt, resume LLM call.
agent_builder.add_edge("interrupt_handler", "llm_call")

# Compile the subgraph.
response_agent = agent_builder.compile()

# --- Part 2: Build the Overall Workflow ---

# Initialize the main graph with input schema `StateInput`.
overall_workflow = (
    StateGraph(State, input=StateInput)
    .add_node("triage_router", triage_router)
    .add_node("triage_interrupt_handler", triage_interrupt_handler)
    .add_node("response_agent", response_agent)
    .add_edge(START, "triage_router")
    .add_edge("triage_router", "triage_interrupt_handler")
    .add_edge("triage_interrupt_handler", "response_agent")
)

# Compile the full workflow graph.
email_assistant = overall_workflow.compile()

In this setup, I have two interrupt nodes:

  • "triage_interrupt_handler" in the main graph

  • "interrupt_handler" in the subgraph

When I call graph.invoke(), I expect the graph to pause execution at the first interrupt (triage_interrupt_handler).
However, instead, the graph triggers both interrupts simultaneously, even though execution should stop after the first interrupt.

Could you help me understand why both interrupts are being scheduled and what I might be doing wrong?

2

Hi @Keshav28291

This is likely how it should be. But for a better understanding of your code, could you provide more of it? Especially what your nodes return, and in the node bodies, I would like to see only the interrupt() invocations, Command and Send invocations too if any.