Theo Browne describes a long-running experiment in which coding agentsAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository. attempted to recreate TypeScript's compiler, type checker and language-server tooling in Rust. He reports that earlier OpenAI-based attempts stalled around 84 percent compatibility, while a later Opus 5.5 effort succeeded by porting Microsoft's Go implementationCode generation uses AI or another automated system to create source code from instructions, examples, schemas, or higher-level specifications. rather than translating the original JavaScript directly. The large headline spending figures are his estimated inference value; he separately estimates substantially lower subscription expenditure and says he has not personally read the resulting code.
Theo Browne distinguishes fast TypeScript-to-JavaScript transpilation from the harder task of reproducing TypeScript's type-checking behavior. His reported comparisons emphasize compatibility, support for Effect's additional diagnostics and performance on real repositories, not speed alone. He cautions that the project inherits upstream bugs and that a faster alternative can still report incorrect errors. These results are presented as his demonstrations and observations, not an independently audited guarantee of a complete drop-in replacement.
Theo Browne's larger motivation is to make development tooling usable within WebAssembly and JavaScript isolates. A prototype combines an in-memory shell and filesystem with compilation and type checking, demonstrating an agent development loop without normal filesystem writes. He argues that execution-environment costs may become more important as inference becomes cheaper, while acknowledging that the shown prototype is not a verified Cloudflare deployment. His account attributes progress less to elaborate prompts than to environment setup, parallel-work coordinationAgent orchestration coordinates AI agents, tools, people, tasks, state, and control flow so a larger workflow reaches a verified outcome., audits and explicit compatibility and performance goals.
Watch on YouTube
