TypeScript

94% of the compilation errors generated by LLMs are type-check failures, a GitHub Octoverse figure quoted in the January 2026 internal report on language fitness for Spec Drive Development. That number does most of the explaining. The report scores languages on an Agent-Language Fit metric (information density per token, hallucination resistance via the type system, feedback loop latency) and puts TypeScript at the optimal balance point: gradual typing lets an agent sketch with permissive types before refining the definitions, and interfaces act as scaffolding that shrinks the model's search space.

The comparison is where the argument sharpens. Rust's borrow checker collides with the probabilistic nature of LLMs and produces correction loops expensive in both tokens and time. Java and C# saturate the context window with getters, setters, and imports, and their training data carries obsolete enterprise patterns that models reproduce. TypeScript's own advantage is partly circumstantial: it dominates the modern web training corpus (Bolt, Lovable, Replit).

Gergely Orosz's September 2025 account of how «Claude Code» was built describes a stack chosen to be "on distribution" for the Claude model, TypeScript alongside React with Ink, Yoga, and Bun. Roughly 90% of that codebase is written by «Claude Code» itself.

The same language carries «toon-format/toon», the serialization format Johann Schopplich published on 22 October 2025, which reports 30 to 60% token savings against JSON and 86.6% retrieval accuracy versus 83.2%.

A language selected for what models already read well, rather than for its design, is a circular criterion nobody in these sources resolves.