AgentRun: Run AI Generated Code Safely

AgentRun is a Python library that makes it easy to run Python code safely from large language models (LLMs) with a single line of code. Built on top of the Docker Python SDK and RestrictedPython, it provides a simple, transparent, and user-friendly API to manage isolated code execution.

AgentRun automatically installs and uninstalls dependencies with optional caching, limits resource consumption, checks code safety, and sets execution timeouts. It has 97% test coverage with full static typing and only two dependencies.

Documentation
Get started in minutes

Why?

Giving code execution ability to LLMs is a massive upgrade. Consider the following user query: what is 12345 * 54321? or even something more ambitious like what is the average daily move of Apple stock during the last week?? With code execution it is possible for LLMs to answer both accurately by executing code.

However, executing untrusted code is dangerous and full of potential footguns. For instance, without proper safeguards, an LLM might generate harmful code like this:

import os
# deletes all files and directories
os.system('rm -rf /')

This package gives code execution ability to any LLM in a single line of code, while preventing and guarding against dangerous code.

Key Features

Safe code execution: AgentRun checks the generated code for dangerous elements before execution
Isolated Environment: Code is executed in a fully isolated docker container
Configurable Resource Management: You can set how much compute resources the code can consume, with sane defaults
Timeouts: Set time limits on how long a script can take to run
Dependency Management: Complete control on what dependencies are allowed to install
Dependency Caching: AgentRun gives you the ability to cache any dependency in advance in the docker container to optimize performance.
Automatic Cleanups: AgentRun cleans any artifacts created by the generated code.
Comes with a REST API: Hate setting up docker? AgentRun comes with already configured docker setup for self-hosting.
Transparent Exception Handling: AgentRun returns the same exact output as running Python in your system - exceptions and tracebacks included. No cryptic docker messages.

If you want to use your own Docker configuration, install this package with pip and simply initialize AgentRun with a running Docker container. Additionally, you can use an already configured Docker Compose setup and API that is ready for self-hosting by cloning this repo.

Unless you are comfortable with Docker, we highly recommend using the REST API with the already configured Docker as a standalone service.

Getting Started

There are two ways to use AgentRun, depending on your needs: with pip for your own Docker setup, or directly as a REST API as a standalone service (recommended).

REST API

Clone the github repository and start immediately with a standalone REST API.

git clone https://github.com/Jonathan-Adly/agentrun
cd agentrun/agentrun-api
cp .env.example .env.dev
docker-compose up -d --build

Then - you have a fully up and running code execution API. Code in --> output out

fetch('http://localhost:8000/v1/run/', {
    method: 'POST',
    headers: {
        'Content-Type': 'application/json'
    },
    body: JSON.stringify({
        code: "print('hello, world!')"
    })
})
.then(response => response.json())
.then(data => console.log(data))
.catch(error => console.error('Error:', error));

Or if you prefer the terminal.

curl -X POST http://localhost:8000/v1/run/ -H "Content-Type: application/json" -d '{"code": "print(\'hello, world!\')"}'

pip install

Install AgentRun with a single command via pip (you will need to configure your own Docker setup):

pip install agentrun

Here is a simple example:

from agentrun import AgentRun

runner = AgentRun(container_name="my_container") # container should be running
code_from_llm = get_code_from_llm(prompt) # "print('hello, world!')"

result = runner.execute_code_in_container(code_from_llm)
print(result)
#> "Hello, world!"

Difference	Python Package	REST API
Docker setup	You set it up	Already setup for you
Installation	Pip	Git clone
Ease of use	Easy	Super Easy
Requirements	A running docker container	Docker installed
Customize	Fully	Partially

Usage

Now, let's see AgentRun in action with something more complicated. We will take advantage of function calling and AgentRun, to have LLMs write and execute code on the fly to solve arbitrary tasks. You can find the full code under docs/examples/

First, we will install the needed packages. We are using mixtral here via groq to keep things fast and with minimal depenencies, but AgentRun works with any LLM out of the box. All what's required is for the LLM to return a code snippet.

FYI: OpenAI assistant tool code_interpreter can execute code. AgentRun is a transparent, open-source version that can work with any LLM.

!pip install groq 
!pip install requests

Next, we will setup a function that executed the code and returns an output. We are using the API here, so make sure to have it running before trying this.

Here is the steps to run the API:

git clone https://github.com/Jonathan-Adly/agentrun
cd agentrun/agentrun-api
cp .env.example .env.dev
docker-compose up -d --build

def execute_python_code(code: str) -> str:
    response = requests.post("http://localhost:8000/v1/run/", json={"code": code})
    output = response.json()["output"]
    return output

Next, we will setup our LLM function calling skeleton code. We need:

An LLM client such Groq or OpenAI or Anthropic (alternatively, you can use litellm as wrapper)
The model you will use
Our code execution tool - that encourages the LLM model to send us python code to execute reliably

from groq import Groq
import json

client = Groq(api_key ="Your API Key")

MODEL = 'mixtral-8x7b-32768'

tools = [
    {
        "type": "function",
        "function": {
            "name": "execute_python_code",
            "description": "Sends a python code snippet to the code execution environment and returns the output. The code execution environment can automatically import any library or package by importing.",
            "parameters": {
                "type": "object",
                "properties": {
                    "code": {
                        "type": "string",
                        "description": "The code snippet to execute. Must be a valid python code. Must use print() to output the result.",
                    },
                },
                "required": ["code"],
            },
        },
    },
]

Next, we will setup a function to call our LLM of choice.

def chat_completion_request(messages, tools=None, tool_choice=None, model=GPT_MODEL):
    try:
        response = client.chat.completions.create(
            model=model,
            messages=messages,
            tools=tools,
            tool_choice=tool_choice,
        )
        return response
    except Exception as e:
        print("Unable to generate ChatCompletion response")
        print(f"Exception: {e}")
        return e

Finally, we will set up a function that takes the user query and returns an answer. Using AgentRun to execute code when the LLM determines code execution is necesary to answer the question

def get_answer(query):
    messages = []
    messages.append(
        {
            "role": "system",
            "content": """Don't make assumptions about what values to plug into functions. Ask for clarification if a user request is ambiguous.\n 
            Use the execute_python_code tool to run code if a question is better solved with code. You can use any package in the code snippet by simply importing. Like `import requests` would work fine.\n
            """,
        }
    )
    messages.append({"role": "user", "content": query})

    chat_response = chat_completion_request(messages, tools=tools)

    message = chat_response.choices[0].message
    # tool call versus content
    if message.tool_calls:
        tool_call = message.tool_calls[0]
        arg = json.loads(tool_call.function.arguments)["code"]
        print(f"Executing code: {arg}")
        answer = execute_python_code(arg)
        # Optional: call an LLM again to turn the answer to a human friendly response
        query = "Help translate the code output to a human friendly response. This was the user query: " + query + " The code output is: " + answer
        answer = get_answer(query)
    else:
        answer = message.content

    return answer

Now let's try it! get_answer("what's the average daily move of Apple stock in the last 3 days?") "The average daily movement of Apple's stock in the last 3 days is approximately $2.60."

How did get this answer?

First, the LLM generated the code to call the Yahoo stock API (via yf) as such:

#AI generated
import yfinance as yf

# Setting the ticker and period for the last 3 days
apple = yf.Ticker('AAPL')
hist = apple.history(period="3d")

# Calculating daily moves (close - open) and their average
moves = hist['Close'] - hist['Open']
average_move = moves.mean()

print(f'{average_move:.2f}')

That code was sent to AgentRun, which outputted: '\r[*********************100%%**********************] 1 of 1 completed\n2.39'

Lastly, the output was sent to the LLM again to make human friendly. Giving us the final answer: $2.39

Customize

AgentRun has sane defaults, but totally customizable. You can change:

dependencies_whitelist - by default any thing that can be pip installed is allowable.
cached_dependencies - these are dependencies that are installed on the image on initialization, and stay there until the image is brought down. [] by default.

It will take longer to initialize the image with cached_dependencies, however subsequent runs using those dependencies would be a lot faster.

cpu_quota - the default is 50000. Here is GPT-4 explaining what does that mean.

In Docker SDK, the cpu_quota parameter is used to limit CPU usage for a container. The value of cpu_quota specifies the amount of CPU time that the container is allowed to use in microseconds per scheduling period. The default scheduling period for Docker is 100 milliseconds (100,000 microseconds).

If you set cpu_quota to 50000, this means that the container is allowed to use 50,000 microseconds of CPU time every 100 milliseconds. Essentially, this limits the container to 50% CPU usage of a single CPU core during each scheduling period. If your system has multiple cores, the container could still potentially use more total CPU resources by spreading the load across multiple cores.

default_timeout - how long is scripts allowed to run for. Default is 20 seconds.
memory_limit - how much memory can execution take. Default is 100mb
memswap_limit - the default is 512mb. Again, here is GPT-4 explaing what memory_mit and memswap do.

In Docker SDK, the memswap_limit parameter is used to control the memory and swap usage of a container. This setting specifies the maximum amount of combined memory and swap space that the container can use. The value is given in bytes.

Here’s how it works:

Memory (RAM): This is the actual physical memory that the container can use.

Swap: This is a portion of the hard drive that is used when the RAM is fully utilized. Using swap allows the system to handle more memory allocation than the physical memory available, but accessing swap is significantly slower than accessing RAM.

You can change any of the defauts when you initalize AgentRun as below.

from agentrun import AgentRun
# container should be running
runner = AgentRun(
container_name="my_container",
# only allowed to pip install requests
dependencies_whitelist = ["requests"], # [] = no dependencies
cached_dependencies = ["requests"],
# 3 minutes timeout
default_timeout = 3 * 60,  
# how much RAM can the script use
memory_limit = "512mb" 
# how much total memory the script can use, using a portion of the hard drive that is used when the RAM is fully utilize
memswap_limit= "1gb" 
) 
code_from_llm = get_code_from_llm(prompt) # "print('hello, world!')"

result = runner.execute_code_in_container(code_from_llm)
print(result)
#> "Hello, world!"

Benchmarks

AgentRun Median execution time is <200ms without dependencies. Dependency installing is usually the bottleneck and depends on the size of package and if the package has many dependencies as well as caching.

Development

To contribute to this library, first checkout the code. Then create a new virtual environment:

cd agentrun
python -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

pip install -e '.[test]'

To run the tests:

pytest

To run the test with coverage

pytest --cov=agentrun tests/

stetelepta/AgentRun