Building a C compiler with a team of parallel Claudes

Abstraction: 16 parallel Claude agents autonomously build Rust C compiler

Key points:

• 16 Claude (Opus 4.6) agents worked in parallel on a shared git repo, each in Docker containers, using a bare-bones synchronization algorithm to avoid duplicate work
• Over ~2,000 Claude Code sessions across two weeks, consuming 2 billion input tokens and 140 million output tokens at ~$20,000 total cost
• The resulting 100,000-line Rust-based C compiler builds Linux 6.9 on x86, ARM, and RISC-V; also compiles QEMU, FFmpeg, SQLite, Postgres, Redis, and Doom
• Key harness design: tests must be nearly perfect verifiers, agents maintain READMEs/progress files, and a CI pipeline prevents regressions; agents run in a bash while true loop
• Parallelism enabled specialization: separate agents handled deduplication, performance optimization, code quality, documentation, and output efficiency
• Key limitation: Opus could not implement 16-bit x86 real-mode code generation within the 32k size limit, so it falls back to GCC for that phase on x86

Connections: Anthropic · Claude Code · AI Agents · Agentic Coding

Source: https://www.anthropic.com/engineering/building-c-compiler