How the Claude Code team thinks about verificationImage: claude.dev
Cooling

How the Claude Code team thinks about verification

Heat 42 · #47 in AI · Cooling2 reports · 2 sources · First reported Oct 15 · Updated Nov 20

The story

About 90% of Claude Code's own code is now written by Claude, according to an Anthropic article by Addy Osmani, Morgan Westlee Lunt and Qing Wang — which raises the question of how the team verifies work it did not write. Between May and September, daily changes grew from 2,500 to more than 30,000, merged pull requests more than doubled from June to September, and the share of reverted changes fell by more than 40%. The team's answer is verification loops: six layers of checks, with human approval scaled to the risk of the change.
Latest developmentNov 18, 12:07
hn_id 50529612 absent from the Round 146 snapshot, so the row leaves the HN top 30

How it unfolded

  1. Nov 18, 12:07
    hn_id 50529612 absent from the Round 146 snapshot, so the row leaves the HN top 30
  2. Oct 15 article re-heating on HN, not a new publication
  3. admitted on HN score alone per Turbo Pascal/Emacs/Alloy precedent
  4. search: only Oct 20-28 interpretations (Medium, notes.junxiong.dev, smartscope, agent-patch-weekly), no November coverage, no current resonance, cross stays 30 per rounds 118-120
  5. verification loops, six layers of checks, human approval scaled by risk
  6. merged PRs more than doubled Jun-Sep, reverted share down >40%, daily changes 2,500 (May) to 30,000+ (Sep)
  7. Oct 15
    Anthropic 2026-10-15 (Addy Osmani, Morgan Westlee Lunt, Qing Wang): ~90% of Claude Code's own code is written by Claude

The angle

How Anthropic's Claude Code team verifies agent-written code: verification loops, six check layers, and human approval scaled by riskSuggested format: Short video · for Developers using AI coding agents

Heat

Now 42 · Peak 42 · 64 recorded checks since Nov 20
Full ranking →