Getting Began with Smolagents: Construct Your First Code Agent in 15 Minutes

Picture by Creator

# Introduction

 
AI has moved from merely chatting with giant language fashions (LLMs) to giving them legs and arms, which permits them to carry out actions within the digital world. These are sometimes referred to as Python AI brokers — autonomous software program packages powered by LLMs that may understand their setting, make choices, use exterior instruments (like APIs or code execution), and take actions to attain particular objectives with out fixed human intervention.

In case you have been eager to experiment with constructing your personal AI agent however felt weighed down by advanced frameworks, you might be in the precise place. In the present day, we’re going to have a look at smolagents, a strong but extremely easy library developed by Hugging Face.

By the top of this text, you’ll perceive what makes smolagents distinctive, and extra importantly, you’ll have a functioning code agent that may fetch dwell information from the web. Let’s discover the implementation.

# Understanding Code Brokers

 
Earlier than we begin coding, let’s perceive the idea. An agent is actually an LLM geared up with instruments. You give the mannequin a purpose (like “get the present climate in London”), and it decides which instruments to make use of to attain that purpose.

What makes the Hugging Face brokers within the smolagents library particular is their method to reasoning. Not like many frameworks that generate JSON or textual content to resolve which software to make use of, smolagents brokers are code brokers. This implies they write Python code snippets to chain collectively their instruments and logic.

That is highly effective as a result of code is exact. It’s the most pure option to categorical advanced directions like loops, conditionals, and information manipulation. As a substitute of the LLM guessing learn how to mix instruments, it merely writes the Python script to do it. As an open-source agent framework, smolagents is clear, light-weight, and excellent for studying the basics.

// Stipulations

To comply with alongside, you have to:

• Python data. You need to be comfy with variables, features, and pip installs.
• A Hugging Face token. Since we’re utilizing the Hugging Face ecosystem, we’ll use their free inference API. You may get a token by signing up at huggingface.co and visiting your settings.
• A Google account is non-compulsory. If you don’t want to put in something regionally, you’ll be able to run this code in a Google Colab pocket book.

# Setting Up Your Atmosphere

 
Let’s get our workspace prepared. Open your terminal or a brand new Colab pocket book and set up the library.

mkdir demo-project
cd demo-project

Subsequent, let’s arrange our safety token. It’s best to retailer this as an setting variable. In case you are utilizing Google Colab, you should utilize the secrets and techniques tab within the left panel so as to add HF_TOKEN after which entry it by way of userdata.get('HF_TOKEN').

# Constructing Your First Agent: The Climate Fetcher

 
For our first mission, we’ll construct an agent that may fetch climate information for a given metropolis. To do that, the agent wants a software. A software is only a perform that the LLM can name. We’ll use a free, public API referred to as wttr.in, which offers climate information in JSON format.

// Putting in and Setting Up

Create a digital setting:

A digital setting isolates your mission’s dependencies out of your system. Now, let’s activate the digital setting.

Home windows:

macOS/Linux:

You will note (env) in your terminal when energetic.

Set up the required packages:

pip set up smolagents requests python-dotenv

We’re putting in smolagents, Hugging Face’s light-weight agent framework for constructing AI brokers with tool-use capabilities; requests, the HTTP library for making API calls; and python-dotenv, which can load setting variables from a .env file.

That’s it — all with only one command. This simplicity is a core a part of the smolagents philosophy.

Determine 1: Putting in smolagents

// Setting Up Your API Token

Create a .env file in your mission root and paste this code. Please change the placeholder along with your precise token:

HF_TOKEN=your_huggingface_token_here

Get your token from huggingface.co/settings/tokens. Your mission construction ought to appear to be this:

Determine 2: Mission construction

// Importing Libraries

Open your demo.py file and paste the next code:

import requests
import os
from smolagents import software, CodeAgent, InferenceClientModel

• requests: For making HTTP calls to the climate API
• os: To securely learn setting variables
• smolagents: Hugging Face’s light-weight agent framework offering:
  • @software: A decorator to outline agent-callable features.
  • CodeAgent: An agent that writes and executes Python code.
  • InferenceClientModel: Connects to Hugging Face’s hosted LLMs.

In smolagents, defining a software is easy. We’ll create a perform that takes a metropolis identify as enter and returns the climate situation. Add the next code to your demo.py file:

@software
def get_weather(metropolis: str) -> str:
    """
    Returns the present climate forecast for a specified metropolis.
    Args:
        metropolis: The identify of the town to get the climate for.
    """
    # Utilizing wttr.by which is a stunning free climate service
    response = requests.get(f"https://wttr.in/{metropolis}?format=%C+%t")
    if response.status_code == 200:
        # The response is apparent textual content like "Partly cloudy +15°C"
        return f"The climate in {metropolis} is: {response.textual content.strip()}"
    else:
        return "Sorry, I could not fetch the climate information."

Let’s break this down:

• We import the software decorator from smolagents. This decorator transforms our common Python perform right into a software that the agent can perceive and use.
• The docstring (""" ... """) within the get_weather perform is crucial. The agent reads this description to know what the software does and learn how to use it.
• Contained in the perform, we make a easy HTTP request to wttr.in, a free climate service that returns plain-text forecasts.
• Kind hints (metropolis: str) inform the agent what inputs to supply.

This can be a excellent instance of software calling in motion. We’re giving the agent a brand new functionality.

// Configuring the LLM

hf_token = os.getenv("HF_TOKEN")
if hf_token is None:
    elevate ValueError("Please set the HF_TOKEN setting variable")

mannequin = InferenceClientModel(
    model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
    token=hf_token
)

The agent wants a mind — a big language mannequin (LLM) that may purpose about duties. Right here we use:

• Qwen2.5-Coder-32B-Instruct: A strong code-focused mannequin hosted on Hugging Face
• HF_TOKEN: Your Hugging Face API token, saved in a .env file for safety

Now, we have to create the agent itself.

agent = CodeAgent(
    instruments=[get_weather],
    mannequin=mannequin,
    add_base_tools=False
)

CodeAgent is a particular agent sort that:

• Writes Python code to unravel issues
• Executes that code in a sandboxed setting
• Can chain a number of software calls collectively

Right here, we’re instantiating a CodeAgent. We move it a listing containing our get_weather software and the mannequin object. The add_base_tools=False argument tells it to not embody any default instruments, preserving our agent easy for now.

// Working the Agent

That is the thrilling half. Let’s give our agent a activity. Run the agent with a particular immediate:

response = agent.run(
    "Are you able to inform me the climate in Paris and in addition in Tokyo?"
)
print(response)

Once you name agent.run(), the agent:

1. Reads your immediate.
2. Causes about what instruments it wants.
3. Generates code that calls get_weather("Paris") and get_weather("Tokyo").
4. Executes the code and returns the outcomes.

Determine 3: smolagents response

Once you run this code, you’ll witness the magic of a Hugging Face agent. The agent receives your request. It sees that it has a software referred to as get_weather. It then writes a small Python script in its “thoughts” (utilizing the LLM) that appears one thing like this:

That is what the agent thinks, not code you write.

weather_paris = get_weather(metropolis="Paris")
weather_tokyo = get_weather(metropolis="Tokyo")
final_answer(f"Right here is the climate: {weather_paris} and {weather_tokyo}")

Determine 4: smolagents remaining response

It executes this code, fetches the information, and returns a pleasant reply. You could have simply constructed a code agent that may browse the online by way of APIs.

// How It Works Behind the Scenes

Determine 5: The interior workings of an AI code agent

// Taking It Additional: Including Extra Instruments

The facility of brokers grows with their toolkit. What if we needed to avoid wasting the climate report back to a file? We are able to create one other software.

@software
def save_to_file(content material: str, filename: str = "weather_report.txt") -> str:
    """
    Saves the offered textual content content material to a file.
    Args:
        content material: The textual content content material to avoid wasting.
        filename: The identify of the file to avoid wasting to (default: weather_report.txt).
    """
    with open(filename, "w") as f:
        f.write(content material)
    return f"Content material efficiently saved to {filename}"

# Re-initialize the agent with each instruments
agent = CodeAgent(
    instruments=[get_weather, save_to_file],
    mannequin=mannequin,
)

agent.run("Get the climate for London and save the report back to a file referred to as london_weather.txt")

Now, your agent can fetch information and work together along with your native file system. This mix of abilities is what makes Python AI brokers so versatile.

# Conclusion

 
In just some minutes and with fewer than 20 traces of core logic, you’ve got constructed a purposeful AI agent. We’ve seen how smolagents simplifies the method of making code brokers that write and execute Python to unravel issues.

The great thing about this open-source agent framework is that it removes the boilerplate, permitting you to concentrate on the enjoyable half: constructing the instruments and defining the duties. You’re now not simply chatting with an AI; you might be collaborating with one that may act. That is only the start. Now you can discover giving your agent entry to the web by way of search APIs, hook it as much as a database, or let it management an online browser.

// References and Studying Assets

Shittu Olumide is a software program engineer and technical author enthusiastic about leveraging cutting-edge applied sciences to craft compelling narratives, with a eager eye for element and a knack for simplifying advanced ideas. You may as well discover Shittu on Twitter.