Building a C compiler with a team of parallel Claudes

Source: Anthropic Engineering Date: 2026-02-05 URL: https://www.anthropic.com/engineering/building-c-compiler

Summary

Nicholas Carlini describes building a 100,000-line Rust C compiler using 16 parallel Claude agents over ~2,000 sessions at roughly $20,000 total cost. Agents claimed tasks via text files, used GCC as a live oracle for parallelizing correctness checks, and relied on high-quality automated tests rather than human supervision. The resulting compiler builds Linux 6.9 on multiple architectures.

Implications

The parallel-agent harness thread. This is Anthropic’s own researcher validating the multi-agent pattern at a non-trivial scale — file-based task claiming as a synchronization primitive is a surprisingly low-tech choice that works precisely because it avoids distributed state. The design is a direct counterpoint to more elaborate orchestration frameworks (LangGraph, CrewAI).

Test-driven autonomy. The explicit framing — “Claude excels when feedback is clear and automated” — is the key harness design insight. High-quality tests are the real orchestrator; the agent loop is secondary. This pressures any shop building Claude Code harnesses to invest heavily in the feedback signal, not the coordination layer.

Cost/scale calibration. $20K for a production-grade compiler-scale project gives a rough upper-bound reference for ambitious autonomous coding tasks. The 16-agent, 2,000-session envelope is now a data point for planning similarly scoped work.