Claude Team Has Claude Write ~90% of Its Code — Here's How They Keep It Working
blaizedsouza · x · 2026-11-17
The Claude Code team says 90% of its code is now written and reviewed by Claude, and shared how they keep that from breaking:
- Verification loops enable running dozens of parallel Claudes instead of 1-2
- A separate grader checking against a rubric raised task success by up to 10 points
- If picking one test type, pick end-to-end: a screenshot of the broken step tells Claude more than "Timeout after 30s"
- Property tests throw thousands of inputs at unbreakable rules like "balance never goes negative"
- Differential tests run old and new systems on the same inputs and flag differences
- A /verify skill in .claude/skills/ boots and drives the real app before every commit
- After 3 failed rounds, Claude stops and calls a human
- A change can never edit the check that judges it; every commit is stamped with how it was verified
Related event: Claude Code team says about 90% of its code is written by Claude(5 posts)→
More from coding & agent
- Developer uses his GitHub Radar project as an agent entry point to decide what to work on next — dSebastien · 2026-11-17
- Lectern CLI turns agent-written Markdown into narrated explainer videos in under 30s — gabe_ragland · 2026-11-17
- Horizon Labs runs an autonomous AI 'company' on one Claude Code subscription, 55k model downloads — xeophon · 2026-11-17
- AI agents narrow formally verified compiler CompCert's gap to gcc -O2 from 28% to 6% — swarat · 2026-11-17
- Expert: Orgs should monitor customer channels for AI incidents like they monitor production logs — simpsoka · 2026-11-17
- IBM's Tejas Kumar on Building Agent Harnesses From Scratch: Guardrails and Daiku — TejasKumar_ · 2026-11-17