The ultimate software engineering abstraction

Really big changes don’t happen in the programming world very often. I’ve been lucky to be around long enough to see a couple of them. And we are all lucky enough to be seeing one today.

I remember well the heady days in the mid-1990s when rapid application development was changing the way software was built. Being able to produce a Windows application with a few drag-n-drops and a few lines of code was a huge leap past the original way of directly accessing the Win32 API to construct a simple window. 

With the release of tools like Visual Basic and Delphi, we finally saw the fruit of object-oriented programming, which allowed us to build applications for Windows much more efficiently. The new IDEs were powerful; I remember the advent of IntelliSense and how cool that was. Clicking on a method declaration and being taken to the implementation was mind-bending. 

But that is all old hat now, eh?

Code what I say

A while back, I wrote about how AI is changing the coding landscape. Then, a couple of months later, I wrote about my experience building a website by “vibe coding” — meaning, using AI to write code interactively, in the flow, based on your intent.

In terms of “AI years,” that was a century ago. We used to talk about “Internet time” and how things were all sped up. But AI? I think things are moving exponentially faster. AI coding is in a kind of virtuous feeding frenzy, where AI is rapidly making AI better. Things are moving so quickly that the model I used back in March is, well, child’s play compared to what the models can do today. 

Of course, “vibe coding” was an informal name. Properly speaking, we are doing “agentic coding,” where a coding agent does all the work, and you just tell it what to do. 

We have IDEs designed specifically for agentic coding. But even the IDE is starting to become less relevant, as OpenAI’s Codex is more like a new kind of command line. You type what you want in plain English, then you review the code changes, and Codex makes a pull request for you. It makes IntelliSense look like some quaint technology of the distant past.  It doesn’t take much imagination to see that this is a huge, order-of-magnitude change in productivity. 

It occurred to me that agentic coding is just a natural step in the continuous application of abstractions. First we had ones and zeros. Those were replaced by assembly language. Then assembly was replaced by higher-order languages. We’ve built things like the JVM and.NET to abstract away the hardware. And of course, the browser pretty much abstracted everything. 

But now? We have created an abstraction layer that can understand people talking and writing. Agentic coding is the compiling of a spoken language into a high-level programming language, be it JavaScript or Ruby or Rust or whatever.

At some point, it won’t matter what the underlying language is. Shoot, let the agent decide what the best tool for the job is. It’s conceivable that agentic coding will advance to the point where the code itself just doesn’t matter.

A higher level

And the really scary part? It’s just getting started. The quality and capabilities of agentic coding are in the infancy stage and only going to get better and better and better. A year ago, we didn’t even think about this kind of thing. Today? If you aren’t on board or getting on board, you will fall behind even faster than if you didn’t pay attention when web development took off. 

Or maybe not. Right now, success with agentic coding requires a certain level of skill. If the agent generates bad code or follows poor practices, it takes a certain amount of expertise to recognize that. It is not far-fetched, however, to imagine the agent becoming so good at coding that expertise will no longer be required.

When high-level languages first required compilers, many thought no machine could write better assembly language than humans. But that concern was put to rest long ago. It’s not unreasonable to think that AI will eventually “compile” your spoken words better than you ever could. Someday, we won’t even look at the code that our AI agent produces. We’ll take it for granted, just as we do the assembly code in our software today.

So maybe the notion of needing special skills to develop software will go away all together. Perhaps coding agents will become so good that anyone with an idea can build an app or a website, all by just talking about it, describing what they want, seeing the results, and iterating. The day is coming when anyone with an idea and a little time can build anything.

And if that isn’t a huge leap in productivity that will change the world, I don’t know what is.