Anthropic used 16 AI agents to build a C compiler in two weeks, a task that took thousands of engineers 37 years.
The researchers ran an experiment to see how far AI could work on its own.
They put 16 AI agents together and asked them to build a full C compiler — a core tool that turns human-written code into computer programs.
Led by safeguards researcher Nicholas Carlini, the AI agents worked in parallel to produce a 100,000-line compiler written in Rust. The tool became capable of compiling the Linux kernel, one of the most widely used operating systems in the world.
With minimal human involvement, the agents divided tasks, checked their own work using tests, and coordinated progress automatically.
The project ran for two weeks, involved nearly 2,000 Claude sessions, and cost about $20,000 in API usage.
The experiment shows AI can now handle large, real-world software projects — while still revealing limits and risks that require human oversight.
Reply With Quote