Sixteen Claude AI agents working together created a new C compiler

In a significant leap for AI collaboration, Anthropic has unveiled an ambitious project where 16 instances of its Claude Opus 4.6 AI model collaborated to construct a C compiler from the ground up. This announcement follows a growing trend towards multi-agent AI systems, with both Anthropic and OpenAI recently launching innovative tools. Nicholas Carlini, a researcher at Anthropic, shared insights into this groundbreaking experiment in a blog post. Tasked with minimal supervision, the AI agents worked on a shared codebase over two weeks, engaging in nearly 2,000 Claude Code sessions that incurred around $20,000 in API costs. Remarkably, they produced a Rust-based compiler spanning 100,000 lines of code, capable of compiling a bootable Linux 6.9 kernel for x86, ARM, and RISC-V architectures. Carlini, who has an extensive background with Google Brain and DeepMind, utilized a new feature from Claude Opus 4.6 known as “agent teams.” Each AI instance operated within its own Docker container, cloning a shared Git repository, claiming tasks through lock files, and submitting completed code back to the main repository without the need for a central orchestration agent. This decentralized approach allowed each instance to autonomously identify and tackle the next most pressing issue. The resulting C compiler, which has been made available on GitHub, has proven capable of compiling various major open-source projects, including PostgreSQL, SQLite, Redis, FFmpeg, and QEMU. It achieved an impressive 99% success rate on the GCC torture test suite and successfully compiled and ran the classic game Doom, which Carlini referred to as “the developer’s ultimate litmus test.” It's important to recognize that creating a C compiler is an ideal task for semi-autonomous AI due to the well-defined specifications and the existence of comprehensive testing suites. In contrast, many real-world software projects lack such clear frameworks, making development challenges more complex. The true difficulty in software engineering often lies not in writing code that passes tests, but in determining what those tests should entail.