The solution
In February 2026 Anthropic researcher Nicholas Carlini described setting 16 Claude Opus 4.6 instances loose on a shared codebase to build a C compiler from scratch. Over two weeks and nearly 2,000 Claude Code sessions, costing about $20,000 in API fees, they produced a 100,000-line Rust compiler that builds a bootable Linux 6.9 kernel on x86, ARM and RISC-V, as Ars Technica reported.
Each instance ran in its own Docker container, claimed tasks by writing lock files and pushed work upstream, with no orchestration agent. Carlini found that verbose test output polluted the context window, so he made test runners print only summary lines. A fast mode sampled 1-10% of tests because Claude has no sense of time.
When all 16 agents got stuck on the same kernel bug, he used GCC as an oracle, compiling most kernel files with GCC and only some with the new compiler so agents worked on different bugs.
What it achieved
Put effort into the environment: quiet test output, sampled tests, and a GCC oracle so each agent fixes a different bug.
Why it worked
Verbose output made agents lose track, so output was kept short.
Sampling tests stopped agents spending hours without progress.
A reference compiler turned one shared bug into many separate tasks.
A well-defined spec, test suites and a reference compiler made the task unusually suited to agents.
What can be applied
With autonomous agents, the harness, verifier and context hygiene decide the outcome as much as the model does.
Aftermath
The compiler passes 99% of the GCC torture test suite and compiled projects such as PostgreSQL, SQLite, Redis, FFmpeg and QEMU, and Doom. Ars Technica lists caveats: it calls GCC for 16-bit x86 boot, has a buggy assembler and linker, produces less efficient code than unoptimized GCC, and degraded around 100,000 lines. It also noted the $20,000 excludes training and human scaffolding, and disputed the 'clean-room' label.
FOLLOW THE EVIDENCE