Why AI agents are so good at coding

I’ve written about how coding is so over. AI is getting smarter every day, and it won’t be long before large language models (LLMs) write better code than any human.

But why is coding the one thing that AI agents seem to excel at? The reasons are simple and straightforward. 

At their core, LLMs process text. They take in massive amounts of text, learn the patterns of that text, and then use all of that information to predict what the next word will be in a given sentence. These models take your question, parse it into text tokens, and then use the trillions (quadrillions?) of vectors they have learned to understand the question and give an answer, one word, or token, at a time. It seems wild, but it is literally that simple. An LLM produces its answer one word at a time.

Doing all this ultimately comes down to just a huge amount of vector math—staggering amounts of calculations. Fortunately, GPUs are really good at vector math, and that is why AI companies have an insatiable appetite for GPUs and why Nvidia is the most valuable company in the world right now. It seems weird to me that the technology used to generate amazing video games is the same that produces amazing text answers to our questions.

Code is text

And of course, code is just words, right? In fact, that is one of the basic tenets of coding—it’s all just text. Git is designed specifically to store and manage text, and to understand the differences between two chunks of text. The tool we all work in, an integrated development environment (IDE), is really a glorified text editor with a bunch of bells and whistles attached. Coding is all about words.

In addition to being words, those words are structured consistently and succinctly—much moreso than the words we speak. Most text is messy, but all code by definition has patterns that are easier for an LLM to recognize than natural language. As a result, LLMs are naturally better at reading and writing code. LLMs can quite quickly and easily parse code, detect patterns, and reproduce those patterns on demand.

Code is plentiful

And there is an enormous amount of code out there. Just think of GitHub alone. A back-of-the-envelope calculation says there is around 100 billion lines of open-source code available for training AI. That’s a lot of code. A whole lot of code. And if you need an explanation of how code works, there are something like 20 million questions and even more answers on Stack Overflow for AI to learn from. There’s a reason that Stack Overflow is a shell of its former self—we all are asking AI for answers instead of our fellow developers.

Code is verifiable

In addition, code is easily verified. First, does it compile? That is always the big first test, and then we can check via testing if it actually does what we want. Unlike other domains, AI’s code output can be checked and verified fairly easily. If you choose to, you can even have your AI write unit and integration tests beforehand, further clarifying and defining what the AI should do. Then, tell your AI to write code that passes the tests. Eventually, AI will figure out that test-driven development is the best path to writing good code and executing on your wishes, and you won’t even have to ask it to do that.

Welcome, Skynet

And finally, code is a great use case for AI agents because developers are generally unafraid of new technology and always seem ready to try out a new tool. This becomes a virtuous circle as AI companies produce coding agents, and developers embrace those coding agents. Software development is a huge part of the economy, and AI companies are strongly incentivized to lean into lucrative markets that are accepting and enthusiastic about using AI agents.

I for one welcome our new coding agent overlords. If there is one area that I’m fine with Skynet taking over, it’s the mundane job of writing structured text that has easily verifiable outcomes. Let the bots grind out code. I’m happy to do the fun part of thinking up and designing new tools and applications.