Human‑in‑the‑Loop Middleware Demo

This repository contains a minimal LangChain/LangGraph example that demonstrates how to pause an agent whenever it wants to call a tool and ask the user for approval (approve / reject).
The pausing logic is handled by HumanInTheLoopMiddleware, which automatically generates a prompt, receives the user’s decision, and resumes execution via a Command(resume={…}).

Why this matters – In many real‑world scenarios an LLM should not act autonomously. By inserting a human checkpoint you can keep control over every tool call, ensuring safety, compliance or simply giving the user a chance to correct mistakes.

📦 Installation

# 1️⃣ Clone the repo
git clone https://github.com/your‑org/human‑in‑the‑loop-demo.git
cd human‑in‑the‑loop-demo

# 2️⃣ Create a virtual environment (recommended)
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

# 3️⃣ Install dependencies
pip install -r requirements.txt

Requirements

Python ≥ 3.10

langchain (>=0.2)

langgraph (>=0.1)

openai or any LLM provider you prefer

The requirements.txt file contains the exact versions used for this demo.

📁 Project Structure

.
├── agent.py          # Agent definition with HumanInTheLoopMiddleware
├── main.py           # CLI entry point – runs the agent interactively
├── tools.py          # Example tool(s) (e.g., get_weather)
└── README.md         # This file

🚀 Running the Demo

1️⃣ Start the Agent

python main.py

You will see a prompt like:

Assistant: What would you like to do?
User: Tell me the weather in London.
Assistant: (Thinking...)

When the agent decides to call get_weather, it pauses and prints:

Human-in-the-loop checkpoint:
The assistant wants to execute tool 'get_weather' with arguments {'location': 'London'}.
Please type one of the following decisions: approve / reject

Type approve to let the tool run, or reject to cancel it.
After your decision, the agent continues:

Assistant: The weather in London is 18°C and sunny.

2️⃣ Using a Different Tool

If you want to test another tool (e.g., get_time), add it to tools.py, import it in agent.py, and adjust the middleware configuration accordingly.

📚 Example Usage

# main.py
from agent import agent

def run():
    print("Welcome! Type 'exit' to quit.")
    while True:
        user_input = input("\nUser: ")
        if user_input.lower() == "exit":
            break
        response = agent.invoke({"input": user_input})
        print(f"Assistant: {response['output']}")

if __name__ == "__main__":
    run()

# agent.py
from langchain.agents import create_agent
from langchain.agents.middleware import HumanInTheLoopMiddleware
from langgraph.checkpoint.memory import MemorySaver

from tools import get_weather  # <-- your tool(s)

memory = MemorySaver()

agent = create_agent(
    model="gpt-4o-mini",          # or any LLM you have access to
    tools=[get_weather],
    system_prompt="You are a helpful assistant.",
    middleware=[
        HumanInTheLoopMiddleware(
            interrupt_on={"get_weather": True},
            description_prefix="Please confirm the tool call:",
        ),
    ],
    checkpointer=memory,
)

# tools.py
from langchain.tools import BaseTool

class GetWeather(BaseTool):
    name = "get_weather"
    description = "Returns current weather for a location."
    args_schema = ...  # define your schema here

    def _run(self, location: str) -> str:
        # Dummy implementation – replace with real API call
        return f"The weather in {location} is 18°C and sunny."

get_weather = GetWeather()

🔧 Customization Tips

Feature	How to change
Allowed decisions	`interrupt_on={"get_weather": {"allowed_decisions": ["approve", "reject"]}}`
Prompt prefix	Change `description_prefix` in the middleware.
Tool list	Add or remove tools in the `tools=[...]` array.
LLM model	Pass a different model name or a custom LLM instance to `create_agent`.

📜 License

This project is licensed under the MIT License – see the LICENSE file for details.

4.3 KiB Raw Blame History Unescape Escape