Use parallel agents to build a compiler

Coordinate many coding agents in containers and back them with tests and an oracle while they work on one compiler repo.

Done with Claude· Advanced

What they did
Nicholas Carlini set up 16 Claude agents as separate processes inside Docker, all contributing to a single shared git repository. The agents used lock files to decide who owned a task, then folded one another’s changes into the codebase without a central manager. Carlini spent most of his effort on the surrounding system: test cases, CI, and a GCC-based reference checker that let different kernel bugs be tackled at the same time.
How it went
The setup ran for about 2,000 sessions and produced a Rust compiler of around 100,000 lines. It can compile Linux 6.9 for x86, ARM, and RISC-V and reaches 99% on most test suites, but it still falls back to GCC for 16-bit x86 and its assembler, linker, and generated code remain problematic. The two-week effort used a little less than $20,000 in API fees.
Tips
A well-designed harness mattered more than agent count; weak tests let the work drift. Keep the whole experiment in a container and expect to spend real time on validation and checks.

Try it yourself with Claude Code

Spin up [number] Claude agents in separate Docker containers to work on [compiler/project] inside one shared git repository. Let each agent reserve a task with a lock file, then have them review and merge each other’s changes as they go. I’ll supply tests, CI, and a reference checker for [target behavior]. Before any irreversible action such as deleting files, sending output, submitting code, or making payments, pause and ask for my approval. Continue iterating on [specific bugs/features] until the project builds and the agreed tests pass, and give short progress updates after each cycle.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.