# A compiler written by 16 Claude agents needed a reliable way to divide its bugs

> Sixteen agents can work at once. The trouble begins when a single broken test gives all of them the same job.

Published 2026-11-08

Canonical: https://machines-on-machines.anth.us/posts/parallel-claude-compiler-tests/

Anthropic researcher Nicholas Carlini asked 16 Claude agents to write a C compiler, the program that translates C source code into instructions a computer can run. The team produced a compiler capable of building Linux 6.9 on several processor architectures, he reports. One difficulty was turning a broken kernel build into separate jobs agents could work on in parallel.

::image{
  src="assets/parallel-claude-compiler-tests/compiler.jpg"
  layout="inline"
  alt="Three green grids show smaller clusters of orange squares from left to right."
  caption="Anthropic’s illustration for its account of the parallel Claude compiler project. The developers used GCC to help narrow which inputs to investigate."
  credit="Anthropic."
}

[Image source](https://www.anthropic.com/engineering/building-c-compiler)

Carlini used GCC, an established compiler, as a comparison tool. His test system compiled most source files with GCC and a selected subset with the agents' compiler. If the combined build failed, the agents could narrow the search to a smaller set of files. Different agents could investigate different parts of the problem.

The experiment took nearly 2,000 Claude Code sessions and about $20,000 in API charges, according to Carlini. The compiler is an inspectable artifact. The testing setup helps explain what made the work possible and how much of the job still fell to the person organizing it.

[Read the source](https://www.anthropic.com/engineering/building-c-compiler).
<!-- papyrus-source-version: 3 -->