After SaaS scare, Anthropic launches new Claude AI with agent teams that build C compilers on their own

Anthropic, the AI startup that shook the stock market a couple of days ago with Claude Cowork, has revealed that its Claude Opus 4.6 model can build a C compiler from scratch. This was a result of an experiment with 16 Claude AI agents working together to accomplish this impressive feat.

On Thursday, Anthropic shared a blog post, detailing this experiment, which was lead by Nicholas Carlini. Carini wrote that 16 parallel Claude agents were tasked to develop a Rust-based C compiler, from scratch, capable of compiling the Linux kernel.

Within two weeks, the agents, running on the Opus 4.6 model, created a 100,000-line compiler, with minimal human intervention. This compiler was created without internet access, and took nearly 2,000 sessions and $20,000 in API costs. The C compiler can build a bootable Linux 6.9 on x86, ARM, and RISC-V.

Claude Opus 4.6 is Anthropic's latest AI model. The model debuts a new “agent teams” feature that allows multiple AI agents to work on a single project simultaneously. Each agent will complete different parts of the project, allowing for better efficiency and productivity.

Why Anthropic’s Claude Opus 4.6 experiment is a big deal

After the AI startup announced this experiment on X, social media was flooded with reactions. Derya Unutmaz, a professor at The Jackson Laboratory, wrote, “You got to be kidding me!”

But why has this left people stunned? A C compiler is a software programme that translates human-readable source code written in the C programming language into machine-readable instructions that a computer's central processing unit (CPU) can understand and execute. This allows computers to receive instructions in binary code (0s and 1s).

This programme is believed to be a highly complex task, and even experienced programmers may struggle when creating a C compiler from scratch. As such, Anthropic’s experiment has left many impressed.

How does the Claude C compiler perform?

Key technical innovations in the Anthropic experiment included a parallel workflow, where each Claude agent operated within its own isolated container, and a simple synchronisation mechanism. This allowed for efficient division of labor and conflict resolution among agents.

The C compiler produced by Claude agents was evaluated against challenging benchmarks.It passed 99 per cent of the GCC torture test suite and was even able to compile and run the classic game Doom. However, the compiler has limitations, including the lack of a 16-bit x86 backend required for Linux booting, reliance on GCC for some stages, and lower efficiency compared to established compilers.

Testing rigour proved crucial. Carlini explained that sustained progress required "extremely high-quality tests" and continuous integration pipelines to ensure that new commits would not break existing code. Writing clear verifier scripts and maintaining up-to-date documentation enabled Claude agents to self-orient and recover from context loss between tasks—a common challenge with current language models.