How to Build a Human-in-the-Loop AI Agent with LangChain & LangGraph

AI agents are powerful — but sometimes too powerful. What if your agent is about to send an email on your behalf and you want a chance to say "wait, hold on"? That's exactly what Human-in-the-Loop (HITL) is for.

In this article, we'll build an AI agent that can send emails, but pauses and asks for your approval before actually doing it. If you approve, it goes through. If you reject, it doesn't. Simple, safe, and production-ready.

---

What We're Building

A LangChain agent powered by an LLM that:

1. Receives a user request like "Send an email to [e-mail]"
2. Decides to call the send_email tool
3. Pauses and waits for human approval
4. Sends the email (or drops it) based on your decision

---

Prerequisites

pip install langchain langgraph

You'll also need an LLM. In this article we'll use OpenAI's GPT-4o-mini via langchain-openai, but you can swap in any LangChain-compatible model.

pip install langchain-openai

Set your API key:

export OPENAI_API_KEY="your-key-here"

---

Step 1: Set Up the Model

from langchain_openai import ChatOpenAI

# You can swap this for any LangChain-compatible chat model
model = ChatOpenAI(model="gpt-4o-mini", temperature=0)

Using a different provider? LangChain supports Anthropic, Mistral, Groq, and many others. Just replace ChatOpenAI with the appropriate class and install its package (e.g. langchain-anthropic for Claude).

---

Step 2: Define the Tool

This is the action the agent will want to take. We keep it simple — just a function that prints the email details.

def send_email(email: str, message: str) -> str:
    """Send email"""

    print(f"\nEMAIL SENT TO: {email}")
    print(f"MESSAGE: {message}")

    return "Email sent successfully"

In a real app, you'd integrate this with SendGrid, SES, or your SMTP server.

---

Step 3: Create the Agent with Human-in-the-Loop Middleware

Here's where the magic happens. We use:

• create_agent — builds the agent with tools attached
• HumanInTheLoopMiddleware — tells the agent to pause before calling specific tools
• InMemorySaver — a checkpointer that saves agent state so it can be resumed after the pause

import warnings

warnings.filterwarnings("ignore", category=UserWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)

from langchain.agents import create_agent
from langchain.agents.middleware import HumanInTheLoopMiddleware
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.types import Command

agent = create_agent(
    model=model,
    tools=[send_email],
    middleware=[HumanInTheLoopMiddleware(interrupt_on={"send_email": True})],
    checkpointer=InMemorySaver(),
)

The key line is interrupt_on={"send_email": True}. This tells the middleware: "Before you actually run send_email, freeze and hand control back to the human."

The InMemorySaver checkpointer is what makes resuming possible — it snapshots the agent's state at the interruption point so you can pick up exactly where you left off.

---

Step 4: First Invocation — Let the Agent Plan

We kick things off with the user's request. The agent will reason about what to do and reach the point of calling send_email — then stop.

config = {"configurable": {"thread_id": "test-thread"}}

response = agent.invoke(
    {
        "messages": [
            {
                "role": "user",
                "content": "Send an email to [e-mail] with message 'hello world'",
            }
        ]
    },
    config=config,
)

print("\nAGENT PAUSED FOR APPROVAL\n")

The thread_id in config is how LangGraph tracks this specific conversation. Use the same thread ID when you resume so it picks up the right checkpoint.

---

Step 5: Ask the Human

Now we prompt the user in the terminal. This is the "human" part of human-in-the-loop.

decision = input("Approve email? (yes/no): ")

---

Step 6: Resume Based on the Decision

This is where Command(resume=...) comes in. We send the agent's paused state a signal telling it whether to proceed or abort.

If Approved

if decision.lower() == "yes":
    response = agent.invoke(
        Command(resume={"decisions": [{"type": "approve"}]}),
        config=config,
    )

    print("\nFINAL RESPONSE:")
    print(response["messages"][-1].content)

The agent resumes from the checkpoint, actually calls send_email, and continues to completion.

If Rejected

else:
    response = agent.invoke(
        Command(
            resume={
                "decisions": [{"type": "reject", "message": "Human rejected the email"}]
            }
        ),
        config=config,
    )

    print("\nEMAIL REJECTED")

The agent is told the action was rejected. It can use the rejection message to inform its final response (e.g., "The email was not sent as per your instruction.").

---

Full Code

Here's everything together:

import warnings

warnings.filterwarnings("ignore", category=UserWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)

from langchain.agents import create_agent
from langchain.agents.middleware import HumanInTheLoopMiddleware
from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.types import Command

# ── Model ──────────────────────────────────────────────────────────────────────
model = ChatOpenAI(model="gpt-4o-mini", temperature=0)


# ── Tool ───────────────────────────────────────────────────────────────────────
def send_email(email: str, message: str) -> str:
    """Send email"""
    print(f"\nEMAIL SENT TO: {email}")
    print(f"MESSAGE: {message}")
    return "Email sent successfully"


# ── Agent ──────────────────────────────────────────────────────────────────────
agent = create_agent(
    model=model,
    tools=[send_email],
    middleware=[HumanInTheLoopMiddleware(interrupt_on={"send_email": True})],
    checkpointer=InMemorySaver(),
)

config = {"configurable": {"thread_id": "test-thread"}}

# ── First call: agent plans and pauses ─────────────────────────────────────────
response = agent.invoke(
    {
        "messages": [
            {
                "role": "user",
                "content": "Send an email to [e-mail] with message 'hello world'",
            }
        ]
    },
    config=config,
)

print("\nAGENT PAUSED FOR APPROVAL\n")

# ── Human decision ─────────────────────────────────────────────────────────────
decision = input("Approve email? (yes/no): ")

if decision.lower() == "yes":
    response = agent.invoke(
        Command(resume={"decisions": [{"type": "approve"}]}),
        config=config,
    )
    print("\nFINAL RESPONSE:")
    print(response["messages"][-1].content)

else:
    response = agent.invoke(
        Command(
            resume={
                "decisions": [{"type": "reject", "message": "Human rejected the email"}]
            }
        ),
        config=config,
    )
    print("\nEMAIL REJECTED")

---

Sample Output

When approved:

AGENT PAUSED FOR APPROVAL

Approve email? (yes/no): yes

EMAIL SENT TO: [e-mail]
MESSAGE: hello world

FINAL RESPONSE:
The email has been sent to [e-mail] with the message "hello world".

When rejected:

AGENT PAUSED FOR APPROVAL

Approve email? (yes/no): no

EMAIL REJECTED

---

Why This Pattern Matters

Human-in-the-loop isn't just a safety net — it's an architectural pattern for building trust between AI and users. Here's when you'd want it:

Use Case	Why HITL Helps
Sending emails / messages	Prevents accidental or hallucinated sends
Database writes	Confirms destructive operations
API calls with side effects	Gives humans a veto before real-world actions
Financial transactions	Compliance and audit requirements
Customer-facing responses	Quality control before publishing

---

Taking It Further

• Async approval: Instead of a terminal input(), send an approval request to Slack, email, or a web dashboard.
• Timeout handling: Auto-reject if no human responds within N minutes.
• Audit logs: Persist every approval/rejection to a database for compliance.
• Multiple tools: Pass multiple tool names to interrupt_on to gate any subset of your agent's capabilities.

---

Wrapping Up

LangGraph's checkpointing + LangChain's middleware make it surprisingly clean to add human oversight to any AI agent. The agent does the thinking; you keep the final say.

If you found this useful, drop a ❤️ and let me know what kinds of agents you're building in the comments!

---