Simon Willison: Rewriting Bun in Rust via AI agents cost $165,000 and is live in Claude Code

Main idea

Jarred Sumner used a sophisticated AI agent architecture to rewrite Bun runtime from Zig to Rust — the project involved coordinated parallel agents, a TypeScript test suite as a conformance benchmark, and adversarial code review. Result: over 1 million lines of LLM-authored code at ~$165,000 in API tokens is now live in production in Claude Code.

Context

Willison highlights this project as an example of 'sophisticated agentic engineering' — not just prompt-and-check, but a fully orchestrated system with parallel agents, adversarial review, and conformance benchmarking. The project is direct evidence that LLM-authored code can be deployed in production-grade runtimes.

Why it matters

This is one of the most significant real-world examples of AI-driven software rewrites with measured costs and a publicly available outcome. It sets a precedent for what is possible with agentic engineering in 2026.

Details / arguments

• Project: Bun rewrite from Zig to Rust
• Method: coordinated parallel AI agents + adversarial code review
• Conformance benchmark: TypeScript test suite
• Cost: ~$165,000 in API tokens
• Result: >1 million lines of LLM-authored code, live in Claude Code
• Published: simonwillison.net, July 8, 2026, 11:57 PM