Building a C compiler with a team of parallel Claudes

Nicholas Carlini set 16 agents working on a shared codebase to write a C compiler in Rust, with no active human intervention while they ran. The result is a 100,000-line compiler that builds the Linux 6.9 kernel, QEMU, FFmpeg, SQLite, PostgreSQL, Redis and Doom, targeting x86, ARM and RISC-V, with a 99% pass rate on most compiler test suites including the GCC torture tests.

The costs are worth recording alongside that. Nearly 2,000 Claude Code sessions across two weeks, just under $20,000, with Opus 4.6 consuming 2 billion input tokens and generating 140 million output tokens.

His four lessons are about making autonomy survivable rather than about the compiler. Write extremely high-quality tests, since with no human watching the tests are the only thing telling an agent whether it is done. Put yourself in Claude’s shoes, and design for what the model can actually see and communicate. Make parallelism easy through task specialisation, so agents are not competing over the same files. And give agents distinct roles with different responsibilities rather than sixteen copies of the same worker.

He is measured about what it means, noting he expects the positive applications to outweigh the negative but that this is a new world. Worth reading next to Cursor’s agent swarms, which attacks the same coordination problem with a planner-and-worker split and a purpose-built version control layer. Note the dogfooding caveat: this is Anthropic writing up what its own models did.