Scarf shifts new development from Haskell to Python over AI workflow concerns

Scarf, the open source usage analytics company founded by Avi Press, is moving new API development from Haskell to Python after seven years of running Haskell in production, citing AI-assisted software development as the reason. No financing, valuation, revenue or customer metrics were disclosed; the significance is technical and cultural, because Scarf has been one of the more visible startup users of Haskell.

Press, who has served on the Haskell Foundation board, said in a post titled “After 7 years in production, Scarf has reluctantly moved away from Haskell” that existing Haskell systems will remain in place while new API routes are written in Python. Over time, he wrote, the Python server is expected to become the main path and Scarf’s Haskell footprint will contract.

His argument is that AI coding agents have changed the cost model for language and build-system choices. Press said Haskell’s compile times and ecosystem friction are a problem when engineers want to run several agents or worktrees in parallel, test different implementations and keep the useful work. “Haskell is in real danger,” he wrote, arguing that teams and ecosystems using AI well will “move much faster” than those that do not.

Why Scarf is changing languages

Haskell, a functional programming language first released in 1990, remains a niche tool. It ranks No. 46 in the TIOBE index, with a rating below 0.5%. Its emphasis on immutability, recursion and strong types has made it popular with a dedicated community, including academics and engineers who value correctness, but it is not a mainstream startup default.

Press has previously defended Haskell for startups despite the hiring difficulty. In a 2023 talk, he said Haskell’s type system made refactoring easier as business requirements changed and allowed documents to be generated from data types. He also said learning Haskell made him a better programmer.

The new claim is that those advantages no longer outweigh the drag in Scarf’s current workflow. Press said Python has already improved some production engineering tasks at Scarf when used with AI tools, including bug fixes that require limited oversight. He described cases where an AI-assisted fix could be completed during a customer call, though he did not provide quantitative productivity data or benchmark the systems publicly.

Press said agents need different things than human developers: quick feedback, easy setup, practical examples and errors that help them repair code. He called for Haskell to improve documentation with copyable production examples, produce more helpful error messages and reduce build times.

Community reaction

The post prompted debate across Hacker News, X, Reddit and Haskell forums. Some Haskell users questioned why Scarf would accept Python’s weaker typing instead of investing in faster Haskell builds. Others rejected the premise that Haskell should change to accommodate large language models.

One Reddit user argued that changing a language to satisfy an AI-related definition of “better” would be misguided, particularly because no one knows how LLM tooling will evolve. On the Haskell forum, another contributor said Press was not taking concerns about AI-related harm seriously. Press also denied an allegation that he had a financial stake in AI companies.

Chris Done, a longtime Haskell programmer based in the UK, wrote that production growth is only one measure of Haskell’s value. He said he would still enjoy the language even if it became irrelevant like Elm or PureScript.

Haskell Foundation Executive Director José Manuel Calderón Trilla urged restraint on X. He said the community’s emphasis on doing things “the right way” should not become a reason to attack people who reach a different conclusion. For Scarf, the decision is narrower: new work is going to Python, while the Haskell systems already in production continue to run.

This story draws on original reporting from The Register.