The Bun project rewrote over one million lines of code from Zig to Rust in just 11 days using Claude AI agents. Despite all tests passing successfully, Zig language creator Andrew Kelley called the result "unreviewed slop," expressing concerns regarding architectural integrity and development culture.

image

What Happened

The Bun team completed a massive migration from Zig to Rust by employing Claude AI agents to automate the process. The refactor covered more than 1,000,000 lines of code and took 11 days, with Claude API costs totaling approximately $165,000. The process demonstrated extreme generation speed—roughly 1,300 lines of code per minute.

Context

Traditionally, such large-scale changes in systems software require months or even years of work from entire engineering teams and multiple cycles of manual code review. Using LLM agents allows for a radical reduction in time and cost, but raises the question of how much automatically passing tests can guarantee the absence of fundamental architectural flaws.

Why It Matters for the Industry

This case establishes a new "Massive AI Refactoring" pattern, allowing for the closure of enormous technical debts in extremely short timeframes. This presents the industry with the challenge of developing new "AI-native" code review standards and automated auditing tools capable of verifying not just syntax and tests, but also architectural purity and adherence to project design systems.

Why It Matters for Users

For developers, this is a signal that AI is already capable of deep refactoring work that was previously considered exclusively human. However, it also serves as a reminder: passing tests automatically is not a guarantee of reliability, and the engineer's role will shift from writing code to designing architectural guardrails and verifying the output of agents.

What Is Not Yet Known / Limitations

The expert community is divided: from skepticism regarding infrastructure reliability to the recognition of new economic opportunities for development.

Sources

Author

Look at AI, Editorial Staff