Anthropic: 16 parallel Claude Opus 4.6 agents build a 100,000-line C compiler that compiles Linux 6.9
Confirmed
The takeaway
It became the reference demo for multi-agent “agent teams” on long software projects.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 10 October 2026
Your AI and this story
- GPT-6 AstraIn its training data
- Claude Opus 5.5In its training data
- Gemini 3.8 FlashIn its training data
- Grok 4.7In its training data
It happened before the cutoffs of all four assistants, so it can be in their training data.
Key facts
- Model: Claude Opus 4.6; 16 Claude instances working in parallel on a shared repository
- Effort: nearly 2 weeks, ~2,000 Claude Code sessions, ~2B input and ~140M output tokens, ~$20,000 in API costs
- Output: ~100,000-line compiler; builds Linux 6.9 on x86, ARM and RISC-V; ~99% pass rate on compiler test suites incl. the GCC torture tests
- Also compiled QEMU, FFmpeg, SQLite, PostgreSQL, Redis and Doom
- Stated limits: no 16-bit x86 code generator, relies on an external assembler/linker, emits less efficient code than GCC, cannot build every project
What happened
Nicholas Carlini (Anthropic Safeguards team) ran 16 Claude Opus 4.6 instances in parallel, each in its own Claude Code loop, with tests as the main coordination signal. The post reports: “The agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.” It also lists the compiler’s limits (no 16-bit real-mode code generator, no own assembler/linker, slower code than GCC).
Why it matters
It was an early concrete data point for weeks-long autonomous multi-agent software work, with cost and token figures published, and it was cited widely in 2026 debates on how far “agent teams” can take large engineering projects.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 primary
Primary
- Anthropic Engineering: Building a C compiler with a team of parallel Claudesanthropic.com, official
Changes
- Filed