Vibe porting and where agentic coding is going

Ive seen a few writeups like this lately. “Vibe porting”.

https://blog.cloudflare.com/vinext/

Essentially cloudflare ported most of nextjs in order to be easier to deploy on their infrastructure. And faster, allegedly.

This quote from a similar project sums up what is really interesting about this for me.

The implementation details didn’t interest me that much (I don’t work on frontend frameworks), but the methodology did. They took the existing Next.js spec and test suite, then pointed AI at it and had it implement code until every test passed.

– https://www.reco.ai/blog/we-rewrote-jsonata-with-ai

LLMs are starting to sit in this fuzzy plane where they let you know less about what it is your asking for because they interpret language into code. Sometimes quite well.

There is a narrative that if you aren’t generating vast amounts of code there is some kind of skill issue, when the reality, as ever, is that coding isn’t really the bottleneck. Specifying is.

It’s an xkcd meme, I think, but sufficiently detailed specification is code itself. In other words, code exists to tell a computer exactly what you want it to do.

These absurdly productive projects are built quickly because the hard work of specification has been done already, in the form of integration tests.

I find the framing “we rebuilt such and such in an afternoon” obnoxious, given the tests and specs took years of thought and design, but it does say something important about the way agentic coding is going.

Basically agents achieving rapid reproduction of existing codebases thanks to them having very thorough test suites and operating in a true TDD fashion.

Essentially, hardcore TDD, of the kind we all read about in books like “extreme programming”, puts specification front and center and gives agents a framework in which to iterate.

The way I see things going is a return to BDD. It seems to me that cucumber.io was built for this moment, and way ahead of it’s time.

I’ll be experimenting with this over the next few weeks, and see if it’s possible to “one shot” something by defining behaviour well up front.