langgraph-dev-human-in-the-loop

Human-in-the-Loop with LangGraph

Enable human intervention at any point in your LangGraph workflow for approval gates, content review, input validation, and feedback loops. Human-in-the-loop is a core LangGraph differentiator that allows indefinite pauses with persistent state, making it ideal for production agents requiring human oversight.

When to Use This Pattern

• Approval gates: Pause before critical actions (API calls, database writes, financial transactions)
• Content review: Allow humans to review and edit LLM-generated content before proceeding
• Tool approval: Review tool calls before execution
• Input validation: Validate user input in multi-turn conversations
• Feedback loops: Incorporate human feedback to guide agent behavior

Core Concepts

Dynamic Interrupts with interrupt()

Dynamic interrupts (recommended for production) pause graph execution from within a node based on the current state. Use the interrupt() function to request human input, which gets surfaced to the client. The graph remains paused until resumed with a value via the Command primitive.

from langgraph.types import interrupt, Command
from langgraph.checkpoint.memory import InMemorySaver

def approval_node(state: State):
    # Pause and request approval
    approved = interrupt({
        "question": "Do you approve this action?",
        "details": state["action_details"]
    })
    return {"approved": approved}

# Compile with checkpointer (required for interrupts)
memory = InMemorySaver()
graph = workflow.compile(checkpointer=memory)

# Initial run - hits interrupt
config = {"configurable": {"thread_id": "thread-1"}}
result = graph.invoke({"action_details": "..."}, config=config)

# Resume with human decision
graph.invoke(Command(resume=True), config=config)

Requirements:

• ✅ Must enable a checkpointer (e.g., InMemorySaver, SqliteSaver)
• ✅ Must provide a thread_id in config
• ✅ Resume with Command(resume=<value>)