Nicholas Carlini, a researcher on Anthropic's Safeguards team, led an ambitious experiment: having a C compiler in Rust built entirely by Claude Opus 4.6 agents working in parallel, with no direct human intervention on the code.
Multi-agent architecture: 16 Claude instances worked simultaneously in isolated Docker containers, sharing an upstream Git repository. Each agent autonomously selects its next task via a file-locking system. Synchronization relies on Git: an agent claims a task, works on it, pushes its changes, then moves to the next one. No explicit communication between agents is required.
Impressive results: in roughly 2,000 Claude Code sessions and $20,000 in API cost (2 billion input tokens, 140 million output tokens), the resulting 100,000-line compiler successfully compiles Linux 6.9 on three architectures (x86, ARM, RISC-V), as well as major projects such as QEMU, FFmpeg, SQLite, PostgreSQL, Redis, and even Doom, with a 99% pass rate on compiler test suites including the GCC torture suite.
Design lessons: the author shares four key takeaways. First, test quality is critical because agents solve exactly what is specified. Second, one must think from Claude's perspective: each agent enters a blank environment with no context, requiring exhaustive documentation and minimal output to preserve the context window. Third, enabling parallelism is essential: easy when many independent tests are failing, but compiling the Linux kernel proved problematic because all agents were fixing the same bugs. The solution: using GCC as a "reference oracle" to randomly distribute the files to compile. Fourth, specializing agents by role (deduplication, optimization, design review, documentation) maximizes efficiency.
Model progression: the article also serves as a capability benchmark. Opus 4 barely produced a functional compiler, Opus 4.5 enabled the first compiler to pass test suites, and Opus 4.6 pushes the boundaries by compiling real large-scale projects.
Security concerns: the author, drawing on a penetration-testing background, notes that passing tests does not guarantee software correctness. Autonomous deployment of code unverified by humans represents a real risk requiring new security strategies. The source code is available on GitHub (anthropics/claudes-c-compiler).