Anthropic released a GitHub-hosted compiler project capable of compiling major open-source projects such as PostgreSQL, SQLite, Redis, FFmpeg, and QEMU, and it achieved a 99 percent pass rate on the GCC torture test suite.

Sixteen Claude Opus 4.6 AI agents worked autonomously in separate Docker containers to collaboratively build a C compiler from a shared codebase using agent teams, without a central orchestrator.

In this experiment, Claude Opus 4.6’s agent teams had each agent clone the same Git repository, claim tasks via lock files, and push code upstream, resolving merge conflicts autonomously without a centralized traffic director.

Carlini notes that C compiler development is a favorable test case for AI coding due to its clear specification, established test suites, and a reference compiler, while many real-world projects lack these advantages and require discovering tests as part of development.

Over about two weeks and roughly 2,000 Claude Code sessions, the effort cost around $20,000 in API fees, yielding a 100,000-line Rust-based compiler capable of building a bootable Linux 6.9 kernel on x86, ARM, and RISC-V.

The compiler successfully compiled and ran Doom, which Carlini describes as the developers’ ultimate litmus test.