AI is changing the world. Don't stay behind. Clear summaries, community insight, delivered without the noise. Subscribe to never miss a beat.

After 7 years in production, Scarf has reluctantly moved away from Haskell

avi.press

July 10, 2026

11 min read

🔥🔥🔥🔥🔥

59/100

Summary

Scarf has transitioned away from Haskell after seven years of production use. The decision reflects challenges faced while using Haskell in their development process.

Key Takeaways

Scarf has moved away from Haskell after seven years of using it in production due to issues with compilation time and ecosystem friction.
The rise of AI and LLMs has changed the economics of software development, making long compile times a significant bottleneck in the development process.
The author emphasizes that while type safety remains important, the cost of typechecking is now more critical in a landscape where AI can generate code quickly.
The need for faster development feedback cycles has led to a preference for languages with shorter build times, especially when using multiple coding agents in parallel.

Read original article

Community Sentiment

Mixed

Positives

The combination of LLMs and Haskell is yielding impressive productivity, especially with type systems that provide vital feedback during development.
Haskell's expressiveness and strong type system seem to help agents converge quickly to functional code, making it a powerful choice despite compile time issues.
Some developers argue that Haskell's compile-time checks make it easier to catch errors, which is crucial when using LLMs for code generation.

Concerns

The slow compile times of Haskell are a significant hurdle, especially for iterative development with LLMs, making it less viable for fast-paced environments.
Critics highlight that relying solely on LLMs could lead to overlooking valuable compile-time checks, undermining the quality of generated code.
Some users feel that the author’s decision to move away from Haskell indicates a troubling shift away from strong type systems that help constrain LLM outputs.