After seven years running production workloads on Haskell, Scarf has migrated new API development to Python. The company's backend powered Servant, Beam, and WAI services handling high-volume package traffic with contractual SLAs. The type system caught real bugs and enforced domain modeling discipline, but compilation times and toolchain friction — Nix, CI, cache management — consumed disproportionate engineering effort.

LLMs fundamentally altered the cost calculus. When agents generate working code in minutes, a 15-minute cold build becomes the dominant latency in the development loop. Parallel agent workflows — spinning up multiple worktrees for concurrent exploration — multiply this tax. Caching helps the best case but fails on deep changes, and the engineering investment to maintain performant caches became its own burden.

Scarf adopted a strangler-fig migration: a new Python API server runs alongside the Haskell one, with traffic routed incrementally. LLMs made porting authentication, database access, and shared models straightforward. Test coverage improved dramatically; hotfixes now ship in minutes. The team spends less time on toolchain maintenance and more on feature delivery.

The broader implication is stark: ecosystems that don't optimize for AI-assisted iteration — cheap, disposable execution contexts and fast cold starts — will lose velocity to those that do. Haskell's compilation model, designed for human-paced development, becomes a liability when the bottleneck shifts from authoring to building.