How Claude Code Rewrote a Million Lines of JavaScript Runtime From Zig to Rust

Technology Programming

Bun is a fast JavaScript runtime built as a competitor to Node.js. Its entire core was originally written in Zig, a systems language designed to be a safer, simpler alternative to C. In early 2026, after Bun's creator joined Anthropic, the runtime was rewritten in Rust — and a large share of that work was done with the help of Anthropic's coding agent, Claude Code.

The scale of the effort is what made it a sensation on developer news this week: roughly one million lines of code were translated across a single pull request, with about 6,755 commits and more than a dozen subagents working in parallel. The result passed 99.8 percent of Bun's existing test suite — a strong signal that the rewritten program behaves the same as the original.

How an AI agent rewrites a program

Large coding agents do not read an entire codebase end to end. Instead, they are fed individual source files or small modules, asked to produce a translation, and the results are merged into a queue for a compiler to verify. Claude Code's subagents fanned out across Zig files in parallel, each emitting Rust translations, while a compiler-driven feedback loop caught and fixed mismatches.

A million lines sounds enormous, but an AI agent handles it by treating the job as thousands of small, identical tasks: read one file, translate it, let the compiler say whether it works. The human's job is no longer to write the code, but to design the pipeline and review the hard parts.

Why Rust instead of Zig

The language switch is not about speed alone. Rust has a much larger ecosystem of libraries, more build-tooling, and a deeper pool of engineers willing to contribute to a systems project. For a runtime that will be maintained for years, Rust's established toolchain and memory-safety guarantees are a practical draw even if Zig is a cleaner design.

The part the agent could not do alone

Two details reveal the limits of the automation. First, the rewritten code contains roughly 13,000 unsafe blocks — Rust's escape hatch that disables its memory-safety checks. Those blocks are legitimate in a low-level runtime, but every one of them has to be justified by a human, because the compiler cannot prove them safe. Second, Claude Code did not write the million lines entirely unsupervised; it operated inside a pipeline that Jarred Sumner, Bun's founder, designed and oversaw.

What this tells us about AI coding

At this scale, AI is a force multiplier for large migrations, not a solo author.
The test suite, not the AI, is the real quality gate — 99.8 percent passed matters.
"Unsafe" zones in systems code remain a human responsibility.
The bottleneck is shifting from writing code to reviewing it.

The Bun rewrite is a landmark because it is one of the first publicly documented cases where an AI agent contributed heavily to a million-line, production-grade systems program. It does not prove that AI can replace systems engineers — it proves that, with the right pipeline and the right safety net of tests, a human can steer an agent across a task that used to take months.