A fast rewrite that turned into a governance debate

Bun creator Jarred Sumner says he used parallel Claude agents to port Bun from Zig to Rust in 11 days, turning a major engineering milestone into a broader argument about AI-generated code, software quality and open-source oversight.

Bun is a JavaScript runtime and toolchain that includes a runtime, package manager, bundler and test runner. In simple terms, a runtime is the environment that executes JavaScript programs, while a toolchain is the set of tools developers use to build, package and test software. Bun gained attention because it aims to be a fast, integrated alternative that works well with the Node.js ecosystem.

Why Bun moved away from Zig

Bun originally used Zig for performance and low-level control. It also chose Apple’s WebKit JavaScriptCore engine rather than Google’s V8, a decision tied to startup speed and memory footprint. As Bun’s user base grew, however, more bugs surfaced. Sumner said the migration was necessary because Bun’s architecture combined garbage collection with application-driven memory management. Garbage collection is an automated way to reclaim unused memory; application-driven memory management means the program itself takes more direct responsibility for memory lifetime.

According to the source material, Anthropic acquired Bun in December 2025 and built its core state machine on top of it. Before that acquisition, a Claude-powered bot called RoboBun was already doing substantial maintenance work in the Bun codebase, including bug fixes and test-failure resolution. It had become the contributor with the most merged pull requests.

Key figures from the rewrite include:

• 11 days to complete the port;
• about 50 dynamic Claude Code workflows running in parallel;
• a peak of roughly 1,300 lines of code per minute;
• more than 1 million lines of Rust generated;
• an estimated $165,000 in API-priced cost;
• Bun’s own test suite of more than 1 million assertions, which Sumner said passed 100% on all supported platforms without skipped or removed tests.

Supporters see a new engineering model

For Sumner, the point was not merely to change languages. He argued that manually rewriting roughly 500,000 lines of Zig would have taken a small engineering team about a year, during which bug fixes, security fixes and feature work would have been disrupted. The AI-assisted approach made a previously impractical rewrite look feasible.

HashiCorp co-founder Mitchell Hashimoto reacted positively on X, arguing that at comparable compensation levels, human engineers could not have reached the same milestone in 11 days. That reaction captures why the story has resonated: AI is no longer being used only for small snippets or autocomplete-style assistance, but for large-scale codebase transformation.

Zig’s creator pushes back

Zig creator Andrew Kelley strongly rejected the idea that Zig itself should be blamed for Bun’s problems. In his view, the issue was not mainly about Rust versus Zig, or even about whether AI was used. It was about two very different value systems in software development.

Kelley said he and others had become increasingly alarmed by programming practices in Bun’s codebase even before Anthropic acquired it. He argued that Bun’s aggressive feature development led to accumulated bugs, weak error handling and significant technical debt. His central objection is that a test suite that failed to catch all bugs in the Zig version should not be assumed to catch every problem in more than 1 million lines of largely unreviewed Rust code.

The dispute also reflects a wider open-source tension around AI-generated contributions. The Bun team previously maintained a Zig branch that reportedly improved debug compile speed by four times, but the Zig project declined the changes under a policy against AI-based contributions. Kelley has argued that large language model submissions often arrive with poor quality and that lack of engineering supervision can create long-term problems.

The bigger lesson: AI makes rewrites possible, not automatically safe

The Bun case does not prove that Rust is simply better than Zig, nor that AI can replace software engineers. Its larger significance is that AI can make large refactors economically and operationally possible, while leaving review, accountability and maintainability unresolved.

Tests are essential, but they only cover the cases engineers have thought to encode. They do not replace architecture review, security auditing or long-term maintenance judgment. For engineering teams, Bun is a useful signal: AI can be powerful for repetitive, mechanical and verifiable migrations, but the closer the code is to critical infrastructure, the more important it becomes to define what was reviewed, what was merely tested and who is responsible for the resulting system.

AI-assisted development will likely continue moving into runtimes, compilers and infrastructure code. At the same time, open-source projects are likely to tighten expectations around provenance, review standards and traceability. Bun’s rewrite may become a landmark case because it shows how fast AI-driven refactoring can be—and why speed alone is not the same as engineering confidence.