Secure AI agent deployment: run Python code for data analysis with LangGraph

Introduction

Today, we’re building a Data Analyst Agent—an AI-powered system that helps you make sense of your data. For example, it can generate visualizations from a CSV file or run complex analyses on structured datasets.

Why does this matter? Large Language Models (LLMs) like ChatGPT already provide powerful data analysis capabilities by generating and executing Python scripts on their servers. While this is convenient, many users are often hesitant to share sensitive data with external platforms. Moreover, you might want to use your own self hosted LLM to reason about your data.

In this article, I’ll show you how you can instruct any LLM to generate and execute code in a secure, ephemeral environment. For this task we are going to use Claude for LLM and Genezio Interpreter tool for running the code in a secured environment.

Deploying an AI-powered data analyst: Key components and execution flow

Our Data Analyst Agent consists of four key components:

1. Frontend: A chat interface where users can upload data files and ask the LLM to analyze them.
2. LangGraph Orchestrator: The agent orchestrates the workflow, managing interactions between components.
3. Genezio Interpreter tool: A LangGraph tool that informs the LLM that it can run Python code to achieve the task. If you want to learn more about Function Invocation and Tooling in the LLM World check this out.
4. Python Executor Environment: Deployed as part of your application, it communicates with the Genezio Interpreter tool to execute the generated code in a temporary, isolated environment to ensure security and privacy.

Python Code Execution: How does it work?

When the AI agent generates Python code, it needs to be executed in a way that ensures security, privacy, and flexibility. Instead of running the code locally inside the LangGraph process, we use the GenezioInterpreter tool, which provides a secured execution environment.

Here’s what happens under the hood:

1. Code Generation: When a request is received, the LLM might decide that to achieve the task it needs to run Python code. It generates the code and passes it to LangGraph, specifying that it wants to use the GenezioInterpreter tool.
2. Execution Request: The GenezioInterpreter tool makes a request to the Python Executor environment deployed on Genezio. All files that the user has uploaded are also passed in this request.
3. Isolated Execution: A short-lived virtual machine (VM) is created to run the code aka the Python Executor. The user’s data and generated code are securely passed to the VM and the code is executed.
4. Result Handling: Once execution is complete, the response is passed back to the GenezioInterpreter tool and the VM is immediately shut down, and all data is deleted.
5. Response Delivery: The GenezioInterpreter tool sends the response to the LLM.

This approach ensures that no data persists beyond execution, making it a great fit for privacy-sensitive applications.

Let’s see how it works.

The LLM was able to answer questions related to the CSV file that I’ve uploaded. You can also play with it here .

Make it your own

It’s time to make this project your own. You can deploy this application in your own account by clicking this button. You will be asked to create a Genezio account and you will also have to set an Anthropic API Key for the Claude LLM. After that you will have your application up and running.

The repository is here . It contains three modules:

• The React application in the client/ folder.
• The Agent server in the server/agent
• The Python Executor in server/genezio_python_executor

About LangGraph

LangGraph is a library within the LangChain ecosystem designed to simplify the development of complex LLM applications. It provides a framework for defining, coordinating, and executing multiple LLM agents (or chains) in a structured manner. By enabling the creation of cyclical graphs, LangGraph facilitates the development of robust, scalable, and flexible multi-agent systems.

About Genezio

Genezio is a cloud platform that allows developers to deploy applications in secure, ephemeral environments. By creating short-lived virtual machines for code execution, Genezio ensures that data remains private and is deleted immediately after execution. This makes it an ideal choice for applications that require secure and temporary code execution environments.

About Claude

Claude is a state-of-the-art Large Language Model (LLM) developed by Anthropic. Designed to generate human-like text, Claude is capable of understanding and generating natural language, making it suitable for a wide range of applications, including data analysis, content creation, and more. Its advanced capabilities enable it to perform complex tasks efficiently and accurately.