You need a lot of linters e.g. to make sure errors are being handled, and the lack of algebraic data types make expressing data difficult.
I do think it has merits, but I’ll take type safety over simplicity
[0]: https://git.sr.ht/~blainsmith/blog/tree/main/item/.build.yml
I'm using CC+Opus 4.7 max effort, and it's produced a working apt cacher from the first phase of development, so far there have only been a few things I've had to ask it to fix. This is over ~52KLOC (counted by "wc -l"), going on day 3 of it working on it. This includes: caching proxy, garbage collection, "http://HTTPS///" kludge (apt-cacher-ng semantics), MITM https proxy, admin website + metrics, deep validation of metadata and rejecting invalid updates, snapshots of upstream state and delayed metadata update until "hot packages" are available after metadata update...
10/10, would go again.
FYI: My agent loop is: "Work on next step, have codex review it, compact", and then a couple rounds at the end of a phase to review the code against the spec, and a couple rounds at the beginning of a phase to create the spec.
In the end, the language doesn't really matter and extremely rarely is the bottleneck.
Python, Java, Go, C#, PHP, Node, Ruby or Scala, for backend stuff, all these languages can do the work 99% of the time. They all have well developed ecosystems, and are sufficiently mainstream to easily find developers.
In the end, these languages are just variants of the same tool, and I find more interesting to just pick one and build something with it, rather than bickering endlessly about which one is better.
In seriousness, Go is a good choice. Or at least it’s not a bad choice, I’d definitely pick go over many other languages. If go had a better type system, it would be damn near perfect.
Though TypeScript's type system is maybe still more powerful - and therefore might have the edge for agents writing code? (Not to mention there's probably more TypeScript in the training data for LLMs, though perhaps there's _better quality_ Go - I'm not a Go dev though so I couldn't comment further on this.)
Having backwards compatibility with 1.0 just makes it easier to maintain software.
The big plus in the modern era is that the simplicity of the language lends to having agents write Go without much fuss. That and the standard library being batteries include lets you direct the agent to use little or no third party dependencies.
Error handling also seems pretty dumb in comparison to Rust. Admittedly Rust is a much more complicated language, but I felt like I could just go learn more Rust instead of bothering with Go and have more fun.
Like, if the gripe is that Go should be better about following XDG conventions, fair enough. But pretty much every language other than C, C++, and Node tends to use system paths for transitives/compilation artifacts/registries/downloads.
Goroutines? Meh. Software transactional memory and green threads? Heck yeah.
An actual type system? Chef's kiss.
Scott Wlaschin from the F# world has written and talked extensively about F# for "boring" software. It works equally well in Haskell. You don't need to use type-level meta programming to spit out a basic service.
Monads are a great honking idea, btw.