After 7 years in production, Scarf has reluctantly moved away from Haskell
Points and comments are a snapshot, not live.
Scarf moved from Haskell to Python after 7 years, citing AI-driven compile-time costs.
Scarf's backend was Haskell for 7 years. The author, a Haskell Foundation board member, says AI changed tradeoffs: long compile times became a bottleneck for parallel agent workflows. Cold builds and cache setup made cheap disposable execution contexts impossible. Scarf started new API work in Python, porting core code with LLM help. They gained faster development cycles and better test coverage. The author argues Haskell must optimize for AI agents: fast bootstrap, clear examples, low friction. Current ecosystem focus on type system research over tooling is misaligned with AI-era needs.
What commenters are saying
Top commenters strongly disagree, arguing expressive type systems are even more valuable with LLMs to catch generated errors. One says LLMs benefit from fast feedback, and Haskell's slow compile kills that advantage; interpreted languages win. Another notes Java also has slow cold compiles but caching mitigates it, questioning the rewrite. Some observe LLMs perform better with constrained type systems like OCaml than TypeScript. A split emerges: those prioritizing type safety and those prioritizing iteration speed. Several point out that build caching can be fixed with less effort than a full rewrite.