What's the best programming language for coding agents?
Choosing a “best” programming language for AI coding agents turns out to be less about raw token efficiency and more about correctness, tooling, and ecosystem conventions. Commenters compare Python, JavaScript, Go, Rust, Ada, C#, Elixir, and more, noting that static typing, strong compilers, and batteries‑included libraries can help agents self‑correct, while highly popular but loosely structured ecosystems may encourage messy or brittle code. The emerging consensus is that modern LLMs perform similarly across many languages, so the most practical choice is often the one with good tooling and clear patterns that the human maintainer already knows well.
Overall takeaway
- Many commenters read the results as: model performance is similar across mainstream languages; language choice matters less than advertised for correctness alone.
- Others argue the benchmarking misses what actually matters in practice: maintainability, tooling, and ecosystem.
Language performance and correctness
- Multiple people highlight that the benchmarks effectively crown Python and JavaScript as best for agents, though some feel the article underplays this.
- Others argue Go, Rust, C#, and C++ work very well with agents, especially when models are told to avoid “dangerous” features or reflective magic.
- There’s pushback on over-interpreting a single study; some note large variance between runs and idiosyncratic ecosystem issues.
Static vs dynamic typing and tooling
- Strong opinion that strict static typing, good compilers, and analyzers (Rust, Go, Ada, OCaml, typed Python, TypeScript, C#) give better feedback loops for agents and reduce silent failures.
- Counterpoint: dynamic languages like Python and JS remain highly effective, especially with type checkers, linters, and test harnesses.
Ecosystem, libraries, and “one way to do it”
- Languages with a “single obvious way” and strong standard libraries (Go, .NET, Django within Python, Dart/Flutter, TypeScript stacks) are seen as particularly agent-friendly.
- “Batteries included” ecosystems and curated libraries reduce the need for agents to pick third-party dependencies, narrowing the search space.
Niche and functional languages
- Mixed experiences with niche or functional languages: some report excellent results in Gleam, Lustre, Odin, Roc, OCaml, Ada; others find models hallucinate features or struggle.
- Training data size matters but is not decisive; language simplicity, stability, and tooling also influence outcomes.
Token efficiency vs real-world cost
- Several comments argue token-compact syntax is a local optimization; most tokens go to reasoning, orchestration, and retries, not raw code size.
- Studies and anecdotes suggest inter-language token differences are modest; architecture, testability, and componentization of the codebase often dominate.
Blog styling tangent
- A sizable subthread debates the article’s minimalist styling: some like the no-bloat aesthetic; others find it hard to read and argue users shouldn’t have to supply their own CSS or rely on reader mode.