Anthropic has published the results of a bold experiment in which 16 instances of the AI model Claude Opus 4.6 worked almost autonomously to create a C compiler. The project was led by Nicholas Carlini, a researcher on Anthropic's Safeguards team, who had the agents work on a shared codebase with minimal human supervision.
Two weeks of work and 100,000 lines of code
Over two weeks and nearly 2,000 Claude Code sessions, the AI agents created a compiler written in Rust containing 100,000 lines of code. Total API costs came to just under $20,000 (approximately CZK 410,000). The resulting compiler can compile a bootable Linux 6.9 kernel for x86, ARM, and RISC-V architectures.
Carlini used a new feature called "agent teams", which was launched alongside the Claude Opus 4.6 model. Each Claude instance ran in its own Docker container, cloned a shared Git repository, reserved tasks using lock files, and uploaded code back after completion. No orchestration agent supervised the work – each instance independently identified the problem that seemed most obvious and began solving it. When code merge conflicts arose, the AI agents resolved them on their own.
What the compiler can do and where its limits lie
The resulting compiler, which Anthropic published on GitHub, can compile a number of major open-source projects, including PostgreSQL, SQLite, Redis, FFmpeg, and QEMU. It achieved a 99% success rate on the GCC torture test suite and, in what Carlini called the "ultimate developer test," compiled and ran the game Doom.
However, Carlini was candid about the compiler's clear limitations. It lacks the 16-bit x86 backend needed to boot Linux from real mode, so it calls GCC for this step. Its custom assembler and linker remain buggy. Even with all optimizations enabled, it produces less efficient code than GCC with all optimizations disabled. Although the Rust code is functional, its quality does not match that of an expert Rust programmer.
"The resulting compiler nearly reached the limits of Opus's capabilities," Carlini wrote. "I tried (really!) to fix several of the limitations mentioned above, but I was not fully successful. New features and bug fixes often broke existing functionality."
The human work behind the automation
Although Anthropic describes the compiler as a "clean-room implementation" because the agents had no internet access during development, this label is somewhat misleading. The base model was trained on a vast amount of publicly available source code, almost certainly including GCC, Clang, and many smaller C compilers.
The CZK 410,000 figure also deserves context. This number covers only the cost of API tokens and does not include the billions spent on training the model, the human labor Carlini invested in building the infrastructure, or the decades of work by compiler engineers who created the test suites and reference implementations.
Carlini devoted considerable effort to building test environments, continuous integration pipelines, and feedback systems tailored to the specific ways language models fail. For example, he found that detailed test output polluted the model's context window, causing it to lose track of what it was doing. He also discovered that Claude has no sense of time and will spend hours running tests without making progress, so he created a fast mode that samples only 1% to 10% of test cases.
Reactions from the developer community and Carlini
Reactions on GitHub were less unequivocal. User mohswell wrote: "If I went to a supermarket, stole a piece of every loaf they had, and mashed it all together, no one would say I made bread from scratch. They would say I was a thief." User Sambit003 remarked: "The longer the AI-generated code I see... the safer I feel. We still have jobs (for many years to come)... just enjoy the hype."
Carlini himself admitted that he felt conflicted about his own results. "Building this compiler was some of the most fun I've had recently, but I did not expect this to be possible at all so early in 2026," he wrote. He also expressed concerns stemming from his previous career in penetration testing, noting that "the idea of programmers deploying software they have never personally verified is a real problem."
The project demonstrates that a year ago, no language model could have created anything resembling a functional multi-architecture compiler, even with this kind of supervision and an unlimited budget. The model hit a ceiling at approximately 100,000 lines of code, suggesting a practical limit for autonomous agentic coding, at least with current models.
Sources: arstechnica.com and theregister.com
