When the automobile was first invented, many looked like—and were indeed called—horseless carriages.

Early websites for newspapers were laid out just like the paper versions. They still are to some extent.

Our computers have “desktops” and “files”—just like an office from the 1950s.

It is even said that the width of our railroad tracks is a direct result of the width between the wheels of a Roman chariot (though that claim is somewhat dubious).

This phenomenon—using new technology in the old technology way—is often called “paving the cow paths.” Cows are not known for understanding that the shortest distance between two points is a straight line, and it doesn’t always make sense to put pavement down where they wear tracks in the fields.

This notion was formalized by the great Peter Drucker, who said “There is surely nothing quite so useless as doing with great efficiency what should not be done at all.” 

All of this got me thinking about AI writing all of our code now. 

Is code necessary?

We developers spend years honing our craft. We read books about clean code, write blogs about the proper way to structure code, and tell ourselves, rightly, that code is meant to be read and maintained as much as it is to be written.

AI coding could change all of that. Your coding agent doesn’t need to see all that great stuff that we humans do. Comments, good variable names, cleanly constructed classes, and the like are all things we do for humans. Shoot, code itself is a human construct, a prop we created to make it easier to reason about the software we design and build. 

I was recently using Claude Code to build an application, and I insisted that he (I can’t help but think of Claude Code as a person) code against interfaces and not implementations, that he design everything with small classes that do one thing, etc. I wanted the code Claude created to be what we developers always shoot for—well-written, easy to maintain, and decoupled. You know the drill. 

And then it occurred to me—are we all merely paving cowpaths? Should agentic AI be concerned with the same things we humans care about when constructing software? Claude wrote comments all over the place—that was for me, not for him. He wrote the code the way that I wanted him to. Does he have a better idea about how to make the software work? 

For that matter, who needs code anyway? It’s not inconceivable that coding agents will eventually just render machine code—i.e., they could compile your native language directly into a binary. (That’s one way to end the language wars!)

Right now we have the process of writing code, reviewing it, compiling it, and running it. We’ve added an extra layer—explaining our intentions to an agent that translates them into code. If that is a cow path—and the more I think about it, the more it does seem a rather indirect way to get from point A to point B—then what will be the most direct way to get from here to there?

The straightest path to software

Every day, our coding agents get better. The better they get, the more we’ll trust them, and the less we’ll need to review their code before committing it. Someday, we might expect, agents will review the code that agents write. What happens to code when humans eventually don’t even read what the agents write anymore? Will code even matter at that point?

Will we write unit tests—or have our agents write unit tests—only for our benefit? Will coding agents even need tests? It’s not hard to imagine a future where agents just test their output automatically, or build things that just work without testing because they can “see” what the outcome of the tests would be. 

Ask yourself this: When is the last time that you checked the output of your compiler? Can you even understand the output of your compiler? Some of you can, sure. But be honest, most of you can’t. 

Maybe AI will come up with a way of designing software based on our human language inputs that is more direct and to the point—a way that we haven’t conceived of yet. Code may stop being the primary representation of software.

Maybe code will become something that, as Peter Drucker put it, should not be done at all.