Sixteen AI Agents Built a C Compiler From Scratch
Source: Slashdot
Overview
Anthropic researcher Nicholas Carlini set 16 instances of Claude Opus 4.6 loose on a shared codebase over two weeks to build a C compiler from scratch. The AI agents produced a 100,000‑line Rust‑based compiler capable of building a bootable Linux 6.9 kernel on x86, ARM, and RISC‑V architectures.
Read the full story on Ars Technica.
Methodology
- The project ran through nearly 2,000 Claude Code sessions and cost about $20,000 in API fees.
- Each Claude instance operated inside its own Docker container, independently claiming tasks via lock files and pushing completed code to a shared Git repository.
- No orchestration agent directed traffic; the agents coordinated solely through the lock‑file mechanism.
- Carlini invested significant effort building test harnesses and feedback systems to keep the agents productive.
Results
- The compiler achieved a 99 % pass rate on the GCC torture test suite.
- It can compile major open‑source projects, including PostgreSQL, SQLite, Redis, FFmpeg, and Doom.
- It successfully builds a bootable Linux 6.9 kernel for x86, ARM, and RISC‑V architectures.
Limitations
- The compiler lacks a 16‑bit x86 backend and falls back to GCC for that step.
- Its assembler and linker remain buggy.
- It produces less efficient code than GCC even when GCC runs with all optimizations disabled.
- The model hit a practical ceiling at around 100,000 lines of code, as bug fixes and new features frequently broke existing functionality.