Skip to main content
Human-in-the-loop (HITL) lets you pause agent execution at specific points to get user input before continuing. This is essential for tool approval, content review, or any workflow that needs human judgment.

​
How it works

  1. Your graph calls interrupt() at a decision point
  2. The run pauses and the thread status becomes "interrupted"
  3. Your application retrieves the interrupt payload (what the agent wants to do)
  4. The user decides what to do (approve, edit, reject, respond)
  5. You resume the run with a command that tells the agent how to proceed

​
Quick example

​
Define a graph with an approval gate

from langgraph.graph import StateGraph
from langgraph.types import Command, interrupt


def human_approval(state):
    """Pause for human approval before executing tools."""
    tool_calls = state["messages"][-1].tool_calls

    # Send interrupt to client with action details
    response = interrupt({
        "action_request": {
            "action": tool_calls[0]["name"],
            "args": tool_calls[0]["args"],
        },
        "config": {
            "allow_accept": True,
            "allow_edit": True,
            "allow_respond": True,
            "allow_ignore": True,
        },
    })

    # Process the human response
    human_response = response[0]

    if human_response["type"] == "accept":
        return Command(goto="tools")
    elif human_response["type"] == "edit":
        # Update tool args with human edits
        state["messages"][-1].tool_calls[0]["args"] = human_response["args"]
        return Command(goto="tools")
    elif human_response["type"] == "response":
        # Human provided a direct response instead of using the tool
        return Command(
            goto="agent",
            update={"messages": [{"type": "human", "content": human_response["args"]}]},
        )
    elif human_response["type"] == "ignore":
        return Command(goto="__end__")

​
Wire it into the graph

builder = StateGraph(State)
builder.add_node("agent", call_model)
builder.add_node("human_approval", human_approval)
builder.add_node("tools", tool_node)

builder.set_entry_point("agent")
builder.add_conditional_edges("agent", should_use_tools, {
    "tools": "human_approval",  # Route through approval instead of directly to tools
    "__end__": "__end__",
})
builder.add_edge("tools", "agent")

graph = builder.compile()

​
Register in aegra.json

{
  "graphs": {
    "agent_hitl": "./src/my_agent/graph.py:graph"
  }
}

​
Client-side flow

​
1. Start a run that triggers an interrupt

import asyncio
from langgraph_sdk import get_client


async def main():
    client = get_client(url="http://localhost:8000")

    thread = await client.threads.create()

    # This will pause when the agent tries to use a tool
    async for chunk in client.runs.stream(
        thread_id=thread["thread_id"],
        assistant_id="agent_hitl",
        input={"messages": [{"type": "human", "content": "Search for AI news"}]},
    ):
        print(chunk)

    # Check thread status
    thread_state = await client.threads.get_state(thread["thread_id"])
    print(f"Status: {thread_state.get('status', 'unknown')}")
    print(f"Interrupts: {thread_state['interrupts']}")


asyncio.run(main())

​
2. Resume with approval

async def approve():
    client = get_client(url="http://localhost:8000")

    thread_id = "your-thread-id"

    # Resume with approval
    async for chunk in client.runs.stream(
        thread_id=thread_id,
        assistant_id="agent_hitl",
        command={"resume": [{"type": "accept", "args": None}]},
    ):
        print(chunk)


asyncio.run(approve())

​
3. Resume with edits

# Edit the tool arguments before executing
async for chunk in client.runs.stream(
    thread_id=thread_id,
    assistant_id="agent_hitl",
    command={"resume": [{"type": "edit", "args": {"query": "AI news 2026"}}]},
):
    print(chunk)

​
4. Respond directly

# Skip the tool entirely and provide a human response
async for chunk in client.runs.stream(
    thread_id=thread_id,
    assistant_id="agent_hitl",
    command={"resume": [{"type": "response", "args": "I already know the answer..."}]},
):
    print(chunk)

​
5. Ignore the tool call

# Skip the tool and end the turn
async for chunk in client.runs.stream(
    thread_id=thread_id,
    assistant_id="agent_hitl",
    command={"resume": [{"type": "ignore", "args": None}]},
):
    print(chunk)

​
Interrupt before/after nodes

You can also set interrupt points without modifying the graph code, using the interrupt_before and interrupt_after parameters on the run: 
# Pause before the "tools" node executes
async for chunk in client.runs.stream(
    thread_id=thread_id,
    assistant_id="agent",
    input={"messages": [{"type": "human", "content": "Search for AI"}]},
    interrupt_before=["tools"],
):
    print(chunk)
Use "*" to interrupt before/after every node: 
interrupt_before="*"  # Pause before every node
interrupt_after="*"   # Pause after every node

​
Checking interrupt status

# Get thread state to see interrupts
state = await client.threads.get_state(thread_id)

if state["interrupts"]:
    for interrupt_data in state["interrupts"]:
        print(f"Interrupt: {interrupt_data}")

​
Response types

TypeDescriptionargs
acceptApprove the action as-isNone
editApprove with modified argumentsModified args dict
responseSkip the action and provide a direct responseResponse string
ignoreSkip the action entirelyNone

​
Important notes

When resuming an interrupted run, use command instead of input. The input and command fields are mutually exclusive — you can’t send both.
  • Interrupts work transparently across subgraph boundaries
  • The thread status changes to "interrupted" when paused and "idle" when completed
  • You can inspect the interrupt payload in thread state to show the user what the agent wants to do
  • Multiple sequential interrupts are supported in a single run