Nicholas Carlini, a researcher at Anthropic, shared an exciting experiment in his latest blog post: 16 Claude Opus 4.6 agents collaborated to write approximately 100,000 lines of Rust code in just two weeks with minimal human intervention, successfully building a fully functional C language compiler.
The achievement was made possible by deploying 16 Claude model instances using the Agent Teams feature. The entire project was conducted with almost no supervision, and the team gradually built the compiler from scratch through about 2000 code sessions. Notably, these AI agents ran in separate Docker containers and collectively cloned a shared Git code repository, forming an efficient collaboration mechanism.
The project cost approximately $20,000 (about RMB 139,000). The final outcome not only included a bootable Linux 6.9 kernel but also supported x86, ARM, and RISC-V architectures. Additionally, the compiler successfully compiled mainstream open-source projects such as PostgreSQL, SQLite, Redis, and FFmpeg. In evaluations using the GCC stress test suite, the compiler achieved a success rate of 99%. Even more surprisingly, it could successfully compile and run the classic game "Doom."
A highlight of this experiment was that there was no dedicated "orchestration agent" in the system. Each AI instance functioned like an independent developer, autonomously identifying urgent tasks, claiming them by writing lock files, and pushing completed code to the upstream. When code merge conflicts occurred, these agents could also negotiate and resolve them autonomously, demonstrating excellent collaborative capabilities.
Anthropic has now open-sourced the project on GitHub, marking a new milestone in AI collaborative programming. This technology not only brings new possibilities to the programming field but also provides valuable practical experience for future AI development.
