If you’re not on the younger side, or you’re not familiar with a lot of the newer Internet tropes and memes, you might think you’re seeing a typo. But the use of something called smolagents to modularize artificial intelligence workflows is actually based on DoggoLingo – that new dialect developed to give voice to cute canines everywhere.
If you do know something about DoggoLingo, you may be familiar with sentences like “I am smol, do me a protec” or “Am smol potat.”
If so, then the name of this new automated code design process will be more familiar to you.
So let’s talk about what these are, and what they do.
A Rubric for AI Impact
First, you have the idea of the AI agent in general. The agent is an AI entity that does things actively instead of passively. There’s the idea that through AI agents, we’re going to give over power to our LLM brethren.
But the Hugging Face survey of what smolagents are also gives you a neat agency “schedule” to show you how this works.
There are five agency levels, with escalating LLM workflow impact.
At level one, the LLM output has no impact on the program flow.
At level two, LLM outputs can determine basic elements of control flow.
At level three, LLM outputs can determine function results.
At level four, LLMs can “control iteration and program continuation.”
Level five is the big one, where one agentic workflow can trigger another agentic workflow, or in other words, two agents can work together.
This is where you start to envision those predictions made by a lot of seasoned experts that eventually, we’ll have entire companies and organizations filled with AI agents instead of humans. The big question, then, is this – can a company full of human workers and leaders compete well with one that is entirely non-human? Can AI sell widgets or provide IT consulting or manufacture cars, better than people can?
To explore this, proponents of smolagents suggest that we use them to implement simplicity and show how components work together.
The writers also describe the interoperability of the smolagent model this way:
“It supports models hosted on the Hub loaded in their transformers version or through our inference API, but also supports models from OpenAI, Anthropic and many others via our LiteLLM integration.”
You can see examples, like a travel planner based on iterative calculations of distances. All of it exemplifies putting these ideas into practice.
Smolagents and the Democratization of Code
Presumably, the end result is that people are going to be able to understand code better.
In some ways, this sort of effort has been going on for awhile. Gilad David Maayan, a writer at Hackernoon, described “code explainers” this way early last year:
“Code explainers are tools designed to make code more comprehensible,” Maayan writes. “They work by breaking down complex lines of code into simpler, understandable chunks. They can highlight the logic behind each code block and depict the flow of data throughout the program. The ability to explain code is essential for software development teams to understand the functionality of the code, carry out debugging, and collaborate efficiently. Automating code explainers takes this a step further. Automated code explainers are advanced tools that use various techniques, including AI and machine learning, to automatically interpret and explain the code. They not only explain the code but also suggest improvements and detect potential errors. This automation can transform and scale up the code review process.”
Anyway, with smolagents, there’s a similar kind of transparency. Observers will see all of these little agents working together, and they’ll be able to visualize a flow chart for what’s happening.
It helps that they’ll also be introduced to these technologies through pop culture references, and terms that have to do with what the average Internet user is already familiar with. If you say something like “I can haz codeprocess?” or put a smiling Shiba Inu on your automation dossier, you’re bringing this obscure innovation to the masses.
And that’s not a smol thing, by any means.
