Ghostty creator Mitchell Hashimoto argues that AI coding agents fail when prioritizing functional code over software structure

@mitchellh if only we had a word for a precise spec of exactly what should happen!

@mitchellh Am I also crazy to think things like cost of editing code base (like literally the dollar amount) can balloon with bad code, too many lines, inflated context, millions of searches, etc?

@mitchellh This so very much. Unless the code is truly intended to be short lived (throwaway) we're just not there yet.

@mitchellh 👏

@mitchellh Do you not think the line will move?

@mitchellh The redpill is that you get 95% of these benefits from simply having proper tooling/using static analysis/performance profiling, and python scripts that just find the stuff this babbles about https://arxiv.org/pdf/2603.24755

@latereigns That just verifies that things don’t have certain classes of bugs, which is a small part of what defines “works”

@mitchellh another banger. ive been very much relating with mitchell lately

@garybernhardt Check your dms Gary!!!

@dpetrou It’ll move. Just the status quo

@mitchellh Most problems I encounter with agents, boil down to context issues. In a human org, the same issues arise but are often fixed as a result of team rituals.

@mitchellh And they're great at doing interfaces for their peers (agents), but not for their owners (humans)

@mitchellh "works" is carrying that whole sentence and it's the one word nobody specified

@mitchellh This reminds me of experiences I've had with offshore dev shops. They'll create exactly what the spec says, no more, no less, and no matter if it actually does what they user expects or not.

@mitchellh People also generally underestimate the complexity of software. A programming language, for instance, looks very simple at surface. You write a function, create a class, update a variable etc. It's usually not a feature, but its combination with everything else that's complex.

@mitchellh The new "it works" using AI is the old "it works on my machine"

@mitchellh True, but I recall spending my entire career making same argument (in my head or out loud) to the human developers who wrote the inefficient spaghetti code the LLMs were trained on. Needless to say, they didn't listen and software quality had already reached a nadir before LLMs.

@mitchellh ngmi

@mitchellh And if you’re specifying that much, you’re basically back to high spec design (waterfall) and we all know that’s bad.

I honestly think synchronous development will continue for a few more years. Need that iteration to arrive at all the tradoffs and nuances.

@mitchellh “If it works who cares what the code looks like” is the same approach that made outsourcing produce terrible code.

Even if you require high test coverage and enforce 500 linter rules outsourced code still usually looked like sausage meat.

I’m loving coding agents for prototypes