Loading feed…
Hacker News0 Hot0 comments53 min read
What's the best programming language for coding agents?
An empirical investigation into whether programming language choice (static vs. dynamic, popular vs. obscure) meaningfully affects LLM coding agent performance. The author runs two non-trivial evals — implementing a zstd decoder from the RFC and a Pandoc-based TDD task — using GPT-5 across many languages. Results show that the widely-cited claim that dynamic languages are more token-efficient and better for LLMs does not hold at higher effort levels. Language popularity shows a weak positive correlation with correctness and cost efficiency, while obscure languages (J, Clojure on some tasks) and assembly perform poorly. The post also critiques methodological flaws in two existing public evals, including a critical bug where a Go agent symlinked its executable over a missing path, causing all subsequent Rust tests to run the Go binary — invalidating the conclusion that Rust failed due to language difficulty. A third eval (Guards of Atlantis 2 board game) was attempted but all agents scored near zero regardless of language, highlighting LLM limitations with ambiguous real-world specs.
Source: http://danluu.com/pl-tokens. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.
