Jarred Sumner ports Bun to Rust in 11 days for ~$165,000, Zig founder calls it “unreviewed slop”

Jarred Sumner says he ported Bun from Zig to Rust in just 11 days, using a fleet of Claude agents running in parallel, and he estimates the project cost about $165,000 at API pricing. The speed is the headline. The bill is the footnote that makes executives lean in.

Sumner also claims the AI-assisted rewrite then cleared Bun's exhaustive test suite of more than one million assertions, passing 100 percent of those tests across all supported platforms without skipping or deleting any. That is the other half of the promise. If “tests passed” truly means “system is safe,” then rewriting massive codebases stops being a multi-year engineering freeze and starts looking like a new operational lever.

So why did Sumner do it now? Because Bun users kept finding bugs as the project grew, and Sumner ties parts of the urgency to a recent code incident: Anthropic's 512,000-line code leak in March. The Register reports that NodeSource attributed that leak to a Bun bundler bug that generated source maps during builds even when told not to. Sumner’s blog post last week frames the Rustification as necessary, and he draws a pretty direct line from architecture mismatch to reliability. Bun’s architecture mixed garbage collection with application-driven memory management. In Sumner’s telling, Zig wasn’t designed for that task, while Rust was “better at automating memory management.”

But the speed of the rewrite did not land with everyone. Andrew Kelley, the creator of Zig, responded with misgivings he said date back before Anthropic’s acquisition of Bun in December 2025. Kelley wrote that the project’s programming practices made him “increasingly horrified,” and he criticized what he described as aggressive feature shipping that piled up bugs, poor error handling, and technical debt. He also argued that Sumner “was already writing slop well before he had access to LLMs.” In other words, the critique is not just about whether AI wrote code. It is about the surrounding engineering system that decides what enters the repo, what gets reviewed, and how quality is enforced when the pace accelerates.

The conflict is less about programming languages and more about value systems. Kelley explicitly downplayed language feature differences and argued that the core issue is “the diverging value systems of the two projects.” He also noted Bun’s prior status in the Zig ecosystem: Bun was one of the largest and highest profile Zig projects, and until Anthropic acquired it, a regular financial contributor to the Zig Software Foundation. In Kelley’s telling, Bun leaving Zig is not a tragedy for Zig. It is a corrective, because Bun was becoming “the publicly presumed poster child for Zig programming language” and, now, no longer. That is a pretty sharp reputational stake for a language founder.

There is a second, operational issue here that boards and platform leaders should treat as a policy question: AI contribution standards. The Zig project would not accept Bun’s upstreamed changes. The reason given was a policy of not accepting AI-based contributions, and The Register reports that Zig had been receiving an influx of LLM-generated submissions of “dubious quality.” Kelley’s argument is that tests can only do so much when code arrives with limited oversight. He points to the scenario where tests might catch some issues in Zig but still fail to protect you when that same unreviewed code gets transplanted into a different runtime and memory model. His core question is blunt: if the tests missed bugs in Zig code, how would they catch bugs across a million lines of unreviewed “slop” in Rust?

Now zoom out to what this means for Anthropic, for enterprises that depend on Bun, and for anyone trying to operationalize AI in software delivery. Bun itself was not a side project. It’s a JavaScript suite: runtime, package manager, bundler, and test runner. Some developers like it as a fast one-stop shop that plays well with Node.js. Under the hood, Sumner used Apple’s low-memory fast-start WebKit JavaScriptCore (JSC) engine rather than Google’s V8 engine. And Anthropic didn’t just buy it as a curiosity. By the time of acquisition, the company built its core state machine on Bun, and the Register reports that a Claude Bot called RoboBun did substantial work in the Bun repo, supplying the most merged PRs of any contributor and fixing bugs and remediating test failures.

That creates a real governance dilemma. In a world where AI agents can generate large volumes of code quickly, “passing tests” can become the de facto quality gate, even if the underlying code review culture is weaker than before. The Sumner story leans into that possibility: 50 dynamic Claude Code workflows, peak generation of about 1,300 lines of code per minute, over a million lines of Rust code generated, 11 days to finish, and then a million-assertion test suite all passing. The Kelley response leans into the counterpoint: throughput can outpace oversight, and tests may not detect everything when the software supply chain shifts from humans reviewing code to systems producing it.

For executives, this is not just a debate between founders. It is a stress test for how quality, risk, and accountability work when the cost of rewriting falls dramatically. If it really can be done in 11 days for roughly $165,000, then teams that assumed deep refactors must be rare may need to rethink their roadmap. Conversely, if “unreviewed slop” can slip through at scale, then “passed tests” may be insufficient as a risk statement to customers, auditors, and internal risk committees. The strategic stake is simple: the industry is moving toward AI-accelerated rewrites, and the real differentiator will be which organizations can pair that speed with engineering oversight that still holds when the logs are messy and the code base is huge.