Anthropic researcher Nicholas Carlini built a 100,000-line Rust-based C compiler capable of compiling the Linux kernel using 16 parallel Claude agents over ~2,000 Claude Code sessions and $20,000 in API costs. The project stress-tests a new "agent teams" architecture where multiple Claude instances collaborate autonomously on a shared codebase without human intervention. The write-up focuses on harness design, test-driven agent guidance, and the limits of long-running autonomous agent teams.
Research
Building a C compiler with a team of parallel Claudes
Anthropic demonstrated autonomous multi-agent collaboration at scale by having 16 parallel Claude instances build a fully-functional C compiler capable of compiling Linux without human intervention.
Saturday, April 4, 2026 12:00 PM UTC2 MIN READSOURCE: Anthropic Engineering BlogBY sys://pipeline
Tags
research
/// RELATED
Safety3d ago
Disneyland Now Uses Face Recognition on Visitors
Disney deploys optional face recognition at theme parks as WIRED rounds up major security stories including NSA testing Anthropic's Mythos for vulnerability detection and healthcare database breaches.
ResearchApr 28
KARL: Mitigating Hallucinations in LLMs via Knowledge-Boundary-Aware Reinforcement Learning
Researchers propose KARL, which combines reinforcement learning with knowledge-boundary awareness to teach LLMs when to decline low-confidence responses, directly tackling the persistent hallucination problem by aligning model outputs with actual training data coverage.