Anthropic's AI Team Builds a Functional C Compiler in Unsupervised Experiment

In a striking demonstration of emerging multi-agent AI capabilities, researchers at Anthropic have successfully orchestrated a team of 16 AI models to create a working C compiler with minimal human oversight. The project, led by research scientist Nicholas Carlini, leveraged the company's newly released 'agent teams' feature for its Claude Opus 4.6 model.

Over a two-week period, the 16 Claude instances operated semi-autonomously within isolated Docker containers, collaborating through a shared Git repository. Without a central director, each AI independently selected tasks, wrote code, and resolved merge conflicts. The effort consumed nearly 2,000 coding sessions, costing an estimated $20,000 in API fees, and produced a 100,000-line compiler written in Rust.

The resulting software is notably robust. It can compile a bootable Linux 6.9 kernel for x86, ARM, and RISC-V architectures and successfully builds major open-source projects like PostgreSQL, SQLite, and Redis. In testing, it passed 99% of the GCC torture test suite and, in a hallmark moment, compiled and ran the classic video game Doom.

Carlini, who previously worked at Google Brain and DeepMind, acknowledged the project's specific advantages. A C compiler has a decades-old, precise specification and extensive existing test suites—conditions rarely found in typical software development. The experiment highlights both the potential and the current boundaries of AI-assisted coding: while powerful for well-defined problems, the more ambiguous work of defining project goals and tests remains a distinctly human challenge. Anthropic has published the compiler's code on GitHub.

Source: Ars Technica