As an experiment, researchers from Anthropic had sixteen AI agents collaboratively build a functional C compiler, without extensive human supervision.
A team of sixteen AI agents has autonomously developed a complete C compiler capable of handling large software projects, including the Linux kernel. The experiment aims to demonstrate the current capabilities of autonomous software development with AI, but at the same time reveals clear limitations and risks.
Functional Compiler
The experiment was conducted at Anthropic, using AI agents powered by Claude Opus 4.6. Multiple instances of Claude were tasked with working in parallel on the same software project. The agents received an assignment and continued to work autonomously, without constant human intervention.
As an advanced demonstration stress test, Anthropic had agents write a C compiler in Rust, entirely from scratch. This compiler ultimately consists of approximately 100,000 lines of code and can compile Linux 6.9 for various hardware platforms, including x86, ARM, and RISC-V.
read also
Claude Opus 4.6 finds 500 vulnerabilities in open source software
The entire process took two weeks, involved nearly 2,000 separate AI sessions, and cost approximately $20,000 in computing power. Anthropic states that an equivalent team of human programmers would need months for a similar task.
Challenge
According to Anthropic, the biggest challenge was not writing the code itself, but designing the environment around the AI. Good tests proved crucial. If tests were unclear or incomplete, the agents solved the wrong problem. Working in parallel also required extra structure, so that agents did not constantly counteract each other.
The result is functional, but not perfect. The compiler is not a full-fledged replacement for existing tools and produces less efficient code. Some components, such as specific startup logic, still require assistance from classic compilers.
(No) Blind Trust
The experiment does show that AI systems can perform complex tasks autonomously. According to Anthropic, this opens up new possibilities for software development, especially for large and long-term projects. At the same time, the experiment is a warning against overly blind trust. Software that no one has fully reviewed can contain errors that only come to light later.
According to the American AI specialist, autonomous agent teams primarily show where the next step lies: not writing code faster, but building systems that can improve themselves in a controlled manner.