The thought crosses my mind that Haskell may be uniquely suited for AI coding using a very small context window (cheap). Haskell encourages small functions and no global state. So you may be able to capture all the relevant context for editing a Haskell function within a few hundred or few thousand tokens. That would be better than some other languages. Plus the strong typing could help AI agents catch errors.
I have not played around with it to see how that plays out with agentic coding. But it does seem like an interesting idea.
As someone who loves Haskell more than any other language, some challenges are
- the tooling is decades behind, say, Rust or Go
- finding the right library in looks very different in Haskell--you frequently start with the signature on Hoogle. Agents can learn this but it's not the same as "web search"
- creating the right solution also looks different. It's usually borne out of thinking about the types and coming up with the correct algebra. Again models can probably learn to create the right types and orient the solution around that, but it's not automatic
- same today as yesterday, laziness is a blessing and a curse. The runtime can do unpredictable things when you suddenly evaluate a deep thunk
- GHC directives effectively mean there are multiple "Haskells"
Some of those are a result of the "avoid success at all costs" mantra. You can't shake that off in a day. It will take a concerted effort to make it more amenable for seamless adoption.
Haskell continues to be my favorite language to write and read, but Rust is the more practical language with a rich type system. If you're looking for something approaching Haskell's expressiveness but with fewer of these issues, check out PureScript.
No way tooling is decades behind. You have a decent LSP, debugger, package manager and REPL.
Laziness is hard to observe, maybe Strict and StrictData would become more popular in use within this context.
I haven't checked in a while now if effects have become the norm in the ecosystem, or if some solution exists for "string" types, but for me all of Haskell's expressivity is lost in the noise of endless conversion function, wrapper types when stacking monads, and import fiddling.
The package manager/build system is anything but decent. Any Haskell project will involve at least 24 hours worth of manual dependency conflict resolution.
No, this is not the reality of using Haskell packages.
The problem you describe was solved more than a decade ago.
You use a Stackage snapshot (https://www.stackage.org/lts) which is a curation of packages that work together, similar to a Linux distribution like Debian, carrying one version per package.
Our company using Haskell has not spent 1 minute doing "dependency resolution" in the last 10 years, not has anybody we know.
> - the tooling is decades behind, say, Rust or Go
If only Rust had something like GHCi.
Stack and Cabal have longer history than cargo, and Stackage for puzzle dependencies.
> - GHC directives effectively mean there are multiple "Haskells"
A bit like macro libraries and what features are enabled where in Rust
> the tooling is decades behind, say, Rust or Go
That's definitely not true, even if that was true maybe 6 years ago. As someone who's uses Haskell daily and also many other languages, I can see Haskell's tooling as more advanced than many others.
I tried building a cross platform shared library. I still couldn't get GHC to generate the library with Haskell's Runtime (RTS) statically linked. Tried multiple AI models and nothing could make it work. Eventually gave up after a month.
Also there's no support for Android and iOS, atleast not without spending months in recompiling GHC with haskell.nix or other third party projects.
Eventually I switched to Rust. It's tooling is so mature that I can focus on developing than compiling GHC.
Despite these demerits, I love Haskell language and lazy eval. I wish I can use it one day.
I built an example project showing how to build shared libraries in Haskell, in 2019.
https://github.com/mietek/haskell-so-example
Think problem is not in advanced features but in stability, integration and streamlining - compared to Rust/cargo.
I think the GHC directives are what hurts Haskell the most. At this point they should just embrace the GHC extensions and make it the "official" Haskell.
Having to enable them in the code is just a hassle. Just make it official and be done with it, just roll it into the language.
Do GHC2021 and GHC2024 not do what you want?
Again, those are _GHC_ extensions, not "standard" Haskell. At this point it's not Haskell, it's Glasgow Haskell.
And those features are not a library that you just throw in, they are compiler specific.
> - the tooling is decades behind, say, Rust or Go
No way. Where vibe-coded Rust contains tons of "unsafe", you can have your vibe-coded Haskell sprinkled with "unsafePerformIO" and "unsafeCoerce" ;)
accursedUnutterablePerformIO or bust :)
> the tooling is decades behind, say, Rust or Go
Can AI not help speed this up?
As someone who DOESN'T use Haskell... What specifically is it missing?
Are you conflating ecosystem with tooling?
> If you're looking for something approaching Haskell's expressiveness but with fewer of these issues, check out PureScript
Rust is quite expressive. Is Haskell really substantially much more?
I do think Rust is a great language for LLMs because I think expressiveness is key.
That's a fair question. It's part ecosystem, part tooling. The ghcup, stack, cabal mishmash is closer to maven than a tool with a modern streamlined UI like cargo.
And yes, Haskell is significantly more expressive than Rust--at the cost of performance.
Haskell is more easily compared with e.g. OCaml than with something like Rust. (It's worth noting that OCaml now has its own 'rustified' development OxCaml.) The main practical difference is that Haskell is idiomatically based on lazy evaluation, whereas OCaml is strict. This makes corecursive patterns a bit easier to express in Haskell, but OCaml and OxCaml are friendlier wrt. performance concerns.
As an intermediate Haskell programmer, I’ve found that using AI to debug or learn Haskell is a better use of time than generating a lot of code with it.
Even without AI most of my Haskell time is spent thinking.
Also, I hand writing Haskell is one of my small after work pleasures.
Much agreed
There are lots of people successfully using agents with Haskell, for example
* https://github.com/digitallyinduced/ihp/ is mainly written with Claude now
* https://jappie.me/haskell-vibes.html is gushing over agent-written Haskell
* https://discourse.haskell.org/t/anti-llm-sentiment-considere... huge thread about the divisiveness of LLM's from some person who loves using them with Haskell
I can see how agent coding is a nice fit for Haskell, since LLM's tend to be best at tasks where you can easily verify the output – and GHC lets you easily verify much more of your domain than say Python (sure you can test some of what dynamic typing doesn't catch, but LLM-written tests do not always do what you think they should[0]). At the same time, the LLMs tend to not code golf unless asked to, and they don't care if they have to wait for dependencies to compile, so that takes away the two major Haskell time sinks.
[0] https://haskellforall.com/2026/05/type-out-the-code#:~:text=...
"no global state" sure but sprawling monad constraints mean you're jumping around more than other languages.
The terseness of Haskell just hides the inherent complexity of the problem. The AI (or human) still needs to uncover that complexity in order to do non-trivial things.
I think you may find helpful this [1] almost 20 years old [2] agent.
The context window it requires for AI coding can be as short as half a dozen of tokens.I suspect that the compact nature of the syntax would require more tokens spent "thinking" to get decent results. It might be more efficient for simple code though. Either way worth testing. Surely someone must've set up a "how well LLMs handle Xlang" benchmark suite.
I haven't seen such a benchmark although maybe it exists.
As far as benchmarks go, I'd also like to see benchmarks that try to find what LLMs are good at. Most of the benchmarks seem designed to give LLMs hard problems and see if they can succeed. In that sense a "good" benchmark is one with a low pass rate.
But if we're going to do agentic coding we also need to know the opposite. We need to know which types of tasks given in which format LLMs will succeed at with like 95%+ accuracy. Then we can more easily build multi prompt pipelines with high confidence in each step.
I think the main reason more time might be spent thinking is because there's relatively less training data on Haskell out in the wild, meaning an agent may have to check back and forth with static analysis to figure out what's valid.
Compact syntax is generally only a good thing for LLMs because it saves context windows and tokens.
Yup, it was mostly pleasure AI coding a company codebase in Scala. I'm considering it for next projects for this reason although my understanding and intuition for Scala code is much weaker than for Python.
I was a little disappointed with the results I was seeing in F#. I thought the .NET + GPT-5.* pull through would be sufficient but it never seemed to work out. Maybe I just wasn’t approaching it correctly.
I am using https://github.com/github/spec-kit and am pretty content. I also tell it in the constitution.md to generate the .fsi files first, then write tests for it before implementation. That might very well be useless superstition since it is soo easy to fool yourself.
Also tell it to prefer easy language constructs with a need to justify anything more complex. I am currently trying to add elmish since its mind numbingly boring explicit plumbing is something I no longer have to do.