Interrupts on LangGraph Platform

LangSmith Product Help
1

Hi,

I’ve been running into issues with interrupts on LangGraph Platform.

Problem Summary

Issue:
LangGraph workflows with repeated interrupt() calls work correctly when running locally but fail to pause for user input on subsequent interrupts when deployed to LangGraph Platform.

Expected Behavior:

  • User provides feedback → workflow processes it → revises plan → shows revised plan → interrupts again for more feedback
  • This cycle should repeat until user explicitly approves

Actual Behavior on Platform:

  • First interrupt works correctly (pauses and waits for user input)
  • After revision, second interrupt call is made but the latest plan, provided during the interrupt by the user, isn’t shown (but the previous one).
  • Logs show that the new plan revision is correctly created (but not shown through the interrupt)

Technical Details

Workflow Structure:

async def suggest_plan(state: DataSummaryState):
    current_plan = state.plan or initial_plan
    answer = interrupt(current_plan) # This should pause every time

    # Process feedback with LLM
    plan_feedback = await llm.with_structured_output(PlanFeedbackReply).ainvoke(...)

    if plan_feedback.plan_accepted:
        return {"plan_accepted": True, ...}
    else:
        return {"plan_accepted": False, "plan_feedback": answer, ...}

# Graph routing: suggest_plan → (if not accepted) → revise_plan → suggest_plan (loop)

Solutions Attempted

  1. Node Structure Refactoring: Split interrupt and processing logic into separate nodes - no improvement
  2. State Management: Added explicit state clearing between iterations - no improvement
  3. Platform-Specific Guards: Added checks for existing feedback to prevent re-execution issues - no improvement

Key Research Findings

  • Local execution works perfectly with the same code
  • LangGraph Platform has different node re-execution behavior after interrupts
  • Platform automatically handles state persistence (no manual checkpointer needed)
  • Issue appears specific to repeated interrupts in iterative workflows on the platform

Current Status

The workflow correctly:

  • Makes the first interrupt call and pauses
  • Processes user feedback
  • Routes to revision logic
  • Returns to the interrupt node
  • Calls interrupt() again

But the second interrupt() does not send the updated state property to the user on LangGraph Platform, while the identical code works locally.

Question

Is there a known difference in how repeated interrupt() calls behave on LangGraph Platform versus local execution? Are there platform-specific configurations or patterns required for iterative human-in-the-loop workflows?