SAN FRANCISCO, February 22, 2026 — A team of artificial intelligence models has successfully built a fully functional C compiler, capable of running a Linux kernel, marking a significant step forward in autonomous software development. The project, spearheaded by researchers at Anthropic, demonstrates the potential for AI to handle complex coding tasks with minimal human oversight.
AI Agents Tackle Complex Coding Challenge
Researchers tasked AI with building a C compiler from scratch, resulting in a functional tool capable of compiling a Linux kernel.
- Anthropic’s Claude Opus 4.6 AI model agents created a 100,000-line Rust-based compiler.
- The compiler can build a bootable Linux 6.9 kernel on x86, ARM, and RISC-V architectures.
- The project cost approximately $20,000 in API fees and took two weeks to complete.
- The resulting compiler successfully compiled and ran the classic game, Doom.
On Thursday, Anthropic researcher Nicholas Carlini published a detailed account of the experiment, outlining how 16 instances of the company’s Claude Opus 4.6 AI model were set to work on a shared codebase. The AI agents were given the ambitious task of constructing a C compiler from the ground up, with remarkably little direction.
Over a period of two weeks and nearly 2,000 sessions utilizing the Claude Code API, the AI agents collaboratively produced a compiler written in Rust. This compiler isn’t just theoretical; it’s capable of building a bootable Linux 6.9 kernel across multiple architectures – x86, ARM, and RISC-V.
AI agents are demonstrating an increasing ability to tackle complex coding tasks with minimal human intervention.
Carlini, who previously spent seven years at Google Brain and DeepMind, leveraged a new feature within Claude Opus 4.6 called “agent teams.” Each AI instance operated within its own Docker container, cloning a shared Git repository. The agents autonomously claimed tasks by creating lock files and then committed their completed code back to the repository. Notably, no central orchestration agent was used to direct the workflow. Each instance independently identified the most pressing issue and began working on a solution. When conflicts arose during code merging, the AI agents resolved them independently.
The resulting compiler, now publicly available on GitHub, has proven its capabilities by successfully compiling a variety of major open-source projects, including PostgreSQL, SQLite, Redis, FFmpeg, and QEMU. It also achieved a 99 percent pass rate on the rigorous GCC torture test suite. As Carlini noted, the ultimate test came when the compiler successfully compiled and ran Doom.
However, it’s important to recognize the specific conditions that made this achievement possible. Building a C compiler is a relatively straightforward task compared to many real-world software projects. The core challenge in most development isn’t simply writing code that passes tests, but rather defining what those tests should be in the first place. The specification for a C compiler is decades old and well-defined, a luxury not often afforded in modern software development.
