Amid a push towards AI brokers, with each Anthropic and OpenAI transport multi-agent instruments this week, Anthropic is greater than prepared to indicate off a few of its extra daring AI coding experiments. However as normal with claims of AI-related achievement, you’ll discover some key caveats forward.
On Thursday, Anthropic researcher Nicholas Carlini revealed a weblog submit describing how he set 16 cases of the corporate’s Claude Opus 4.6 AI mannequin free on a shared codebase with minimal supervision, tasking them with constructing a C compiler from scratch.
Over two weeks and almost 2,000 Claude Code periods costing about $20,000 in API charges, the AI mannequin brokers reportedly produced a 100,000-line Rust-based compiler able to constructing a bootable Linux 6.9 kernel on x86, ARM, and RISC-V architectures.
Carlini, a analysis scientist on Anthropic’s Safeguards crew who beforehand spent seven years at Google Mind and DeepMind, used a brand new characteristic launched with Claude Opus 4.6 referred to as “agent groups.” In apply, every Claude occasion ran inside its personal Docker container, cloning a shared Git repository, claiming duties by writing lock information, then pushing accomplished code again upstream. No orchestration agent directed visitors. Every occasion independently recognized no matter drawback appeared most evident to work on subsequent and began fixing it. When merge conflicts arose, the AI mannequin cases resolved them on their very own.
The ensuing compiler, which Anthropic has launched on GitHub, can compile a variety of main open supply tasks, together with PostgreSQL, SQLite, Redis, FFmpeg, and QEMU. It achieved a 99 % cross fee on the GCC torture take a look at suite and, in what Carlini referred to as “the developer’s final litmus take a look at,” compiled and ran Doom.
It’s value noting {that a} C compiler is a near-ideal process for semi-autonomous AI mannequin coding: The specification is a long time previous and well-defined, complete take a look at suites exist already, and there’s a known-good reference compiler to test in opposition to. Most real-world software program tasks have none of those benefits. The exhausting a part of most improvement isn’t writing code that passes checks; it’s determining what the checks needs to be within the first place.
