With AI agents being able to write code faster than we can type, writing code is no longer the long pole in software development. That’s the same first sentence I used in my last journal entry. With AI agents orchestrating software development more, the specification is becoming the new source code–and tests are the coded specification.
SQLite is semi-famous for its testing. “The project has 590 times as much test code and test scripts” as the source code. Tests are what enable you to modify code without worrying about breaking things, and when you have that many tests, it becomes pretty trivial to rewrite in the best language or with the best optimizations possible.
We’re increasingly seeing Rust being used as the AI agent language of choice because of safety and the lack of the overhead of a garbage collector. When humans aren’t writing the code, “convenience” and “ease of use” are not features you really need in a programming language. Maybe we’ll see a lower level safe language in the future that works better than Rust, and the important thing is that we have the tests and oracles to be able to safely rewrite anything into that better language.
A lot of the free software world is getting caught up with all the AI stuff and fretting about the copyright implications of AI generated code. These days, I find myself beginning to view source code itself as looking at disassembly or some intermediate language (i.e. unimportant). It tells me how a program will execute, sure, but the specification is really in the tests. If it’s unspecified (no test coverage), you could call that “undefined behavior” the same way as a C program is compiled by a compiler and decides to do something you might not like. Looking ahead, I would imagine more emphasis to be placed on tests rather than the code.