Building a C compiler with a team of parallel Claudes
Sixteen parallel Opus 4.6 agents wrote a 100k-line Rust C compiler in two weeks (~$20k) that builds Linux 6.9 — the writeup credits test quality and context hygiene, not raw model capability.
**The gist** Anthropic pointed **16 parallel Opus 4.6 agents** at writing a C compiler in Rust: **100,000 lines** that compile **Linux kernel 6.9** on x86, ARM, and RISC-V, plus QEMU, SQLite, and PostgreSQL, passing 99% of GCC's torture suite. Roughly 2,000 sessions over two weeks cost about **$20,000**.
**Why it matters** The transferable part is environment design: agents coordinated through a shared **bare git repo** with **lock files** to claim tasks, logs went to files instead of context, and each agent ran a random **1–10% test sample** to avoid context pollution. Near-flawless tests mattered more than the model — Claude solves whatever the harness rewards.
**The gist** Anthropic pointed **16 parallel Opus 4.6 agents** at writing a C compiler in Rust: **100,000 lines** that compile **Linux kernel 6.9** on x86, ARM, and RISC-V, plus QEMU, SQLite, and PostgreSQL, passing 99% of GCC's torture suite. Roughly 2,000 sessions over two weeks cost about **$20,000**. **Why it matters** The transferable part is environment design: agents coordinated through a shared **bare git repo** with **lock files** to claim tasks, logs went to files instead of context, and each agent ran a random **1–10% test sample** to avoid context pollution. Near-flawless tests mattered more than the model — Claude solves whatever the harness rewards. **Watch out** Generated code runs slower than **GCC with optimizations off**, the assembler and linker remain buggy, and late improvements kept introducing regressions — a sign Opus 4.6 is near its practical limit here. The author is candid about the risk of shipping code **no human personally verified**.