Claude Code team: ~90% of code written by Claude, verification loops unlock massive parallelism
addyosmani · x · 2026-11-17
Addy Osmani of the Claude Code team explains how they think about verification.
- Roughly 90% of the team's code is written and reviewed by Claude, making evidence-based verification critical at scale
- Verification loops are the building block for delegation: without them you can only run one or two Claudes at a time; with them, dozens or hundreds in parallel
- Internal testing on structured file generation showed a separate grader checking outputs against a rubric improved task success by up to 10 points over standard prompting, with the largest gains on the hardest problems
- The post highlights three tests worth investing in: E2E, property-based, and differential testing, plus end-to-end tests that provide traces, screenshots, and failure-step descriptions instead of blind unit-test timeouts
Related event: Claude Code Team Says About 90% of Its Code Is Written by Claude(2 posts)→
More from coding & agent
- A $200/month one-person AI company in Claude Code: 6 agent files, 1 human approver — shivsingh · 2026-11-17
- Talk on time-travel testing in a chaotic multiverse for autonomous testing now live — moarbugs · 2026-11-17
- Open-source Ardia browser plays 4K Netflix in Chromium via FairPlay, beats rivals' agents on token use — itsOmSarraf_ · 2026-11-17
- Dev builds Hyo, a personal AI agent plugin for Obsidian, over 6 months — evielync · 2026-11-17
- Built with ChatGPT: Forge extension tames 30,000-LoRA collections with incremental loading — Artistic-Account-423 · 2026-11-17
- Open-source PromptPacker runs Claude Code and Codex directly inside Google Colab — A_Watermelon_Add · 2026-11-17