Building a C compiler with a team of parallel Claudes
- Source
- anthropic.com
- Author
- Anthropic Engineering
- Date
Context
Parallel coding agents helped only once their failures could be separated into independent tasks. In Anthropic's C compiler experiment, sixteen agents initially collided while trying to fix the same Linux build failure, overwriting one another's changes. The researcher changed the tests to compile most files with the established GCC compiler and selected files with the experimental compiler. That helped isolate faults so agents could repair different files.
The team eventually produced a compiler capable of building Linux 6.9, according to the report. Significant limitations remained, including less efficient generated code than GCC with optimizations disabled and reliance on GCC for the 16-bit part of the x86 boot process. The experiment shows how test design can enable parallel progress on a difficult project.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
- multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
“To stress test it, I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel.”
“Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.”
“Merge conflicts are frequent, but Claude is smart enough to figure that out.”
“In most cases, Claude picks up the “next most obvious” problem.”
Checking sign-in…
Loading comments…