Anthropic researcher Nicolas Carlini tried to get the newly releasedOpus 4.6 model to produce a C compiler. The experiment left him with mixed feelings of "excitement," "concern," and "embarrassment."
Anthropic researcher Nicolas Carlini tried to get the newly releasedOpus 4.6 model to produce a C compiler. The experiment left him with mixed feelings of "excitement," "concern," and "embarrassment."
Carlini tasked 16 agents with writing a C compiler from scratch in Rust that could compile the Linux kernel. After nearly 2,000 Claude Code sessions and spending $20,000 on the API, the team of agents created a 100,000-line compiler that could compile Linux 6.9 for x86, ARM, and RISC-V architectures, The Register reports .
According to him, the work of agent teams looks like this: “multiple instances of Claude work in parallel on a common code without active human intervention.”
One of the key challenges was to get around the need for an “operator who needs to be online and ready to collaborate.” This is likely to mean that Claude Code no longer has to wait for human guidance for each next step.
“To achieve sustained autonomous progress, I created a shell that loops Claude in a simple algorithm… As soon as he completes one task, he instantly takes on the next one,” the researcher explained.
Carlini let each Claude agent decide for itself what to do. In most cases, Claude would choose the next most obvious problem. This taught him a number of lessons, including the need to “write extremely high-quality tests.”
The researcher advised “putting yourself in Claude’s shoes.” This means that “the test shell should not output thousands of unnecessary bytes” so that Claude can more easily find the information he needs.
He also added: "Claude has no sense of time and, if left unattended, will happily spend hours running tests instead of moving forward."
Over nearly 2,000 Claude Code sessions in two weeks, Opus 4.6 consumed 2 billion input tokens and generated 140 million output tokens; the total cost was just under $20,000.
This made the project “extremely expensive” compared to even Claude Max’s most expensive plans, Carlini noted. “But that’s a fraction of what it would have cost me to build it myself, let alone hire a whole team.”
The compiler successfully compiles many projects, but not all. It is not yet a full replacement for a real compiler. Moreover, the generated code is not very efficient.
The quality of Rust code is acceptable, but far from the level that an experienced Rust developer could produce.
Carlini concluded: “Agent teams demonstrate the ability to autonomously implement complete complex projects.”
But as a former pentester, he noted that fully autonomous development carries real risks. “The idea that programmers would be implementing software that they’ve never personally tested is a serious concern.” Ultimately, the experiment “excited me, but it also left me feeling uneasy.”
