Anthropic Barely Reads Claude's Code Anymore

Highlights

• Claude writes approximately 90% of Claude Code's own code.

• Anthropic's main monorepo grew from roughly 2,500 changes per day in May to over 30,000 on its busiest day in September.

• Between June and September, merged pull requests more than doubled while the reverted share dropped by 40%.

• Each automated code review using Anthropic's reviewer averages around $15 to $25 in token costs.

• An automated review loop without restrictions once caused an agent-authored PR to undergo over 50 rounds of review overnight.

• Limiting reviews to three rounds increased the share of PRs passing after one round from 65% to 78% and decreased PRs needing five or more rounds from 9% to 1.5%.

Timeline

1. 00:00 - 02:21

   Scaling Code Generation and Review Volume

   Rapid development speeds make manual line-by-line reading unrealistic. Instead of traditional bottlenecked reviews, the process aims for a maximum of one human involvement per pull request, measured against production error rates and revert frequencies to ensure reliability does not degrade.

   • Claude generates about 90% of the code for Claude Code itself.
   • Monorepo activity escalated from 2,500 daily changes in May to over 30,000 on peak days in September.
   • Pull request review strategy shifted from checking if a human read every line to identifying specific failure risks and automated checks.
2. 02:21 - 03:36

   Automated Review Layers and Configuration Files

   A multi-layered defense model catches errors progressively. If reviewers repeatedly flag systemic mistakes, teams introduce permanent lint rules or update guideline files. Specialized review agents independently scan codebases for security vulnerabilities before standard CI checks execute.

   • Pre-PR checks run inside engineer sessions using sub-agents and test execution.
   • Automated AI reviewer agents vote on identified bugs before any human sees the pull request.
   • Review.md files instruct reviewers on what actual failures look like in specific codebase areas.
3. 03:36 - 06:00

   Risk Tiers, Evals, and Managing Review Loops

   Low-risk changes like documentation bypass human approval entirely when submitted by trusted authors, while high-risk paths require two human approvals. Tracking agent noise prevents agents from wasting cycles on old code, significantly increasing first-round merge success rates.

   • A classifier agent categorizes pull requests into risk tiers to dictate human review requirements.
   • Daily evals test review agents against historical incident data to maintain detection accuracy.
   • Limiting review loops to three iterations prevents agent authors and agent reviewers from getting trapped in endless cycles.
4. 06:00 - 07:43

   Human Responsibilities and Economic Realities

   Humans focus on architectural decisions, high-risk billing logic, and final accountability. Although extensive token usage requires substantial resources, smaller teams can still adopt core practices like short configuration files, explicit review guidelines, and strict iteration caps.

   • Human engineers retain ownership of outcomes under the principle of landing and owning code.
   • Token costs average $15 to $25 per automated review pass.
   • While per-PR revert shares dropped, total absolute reverts increased because total merged pull request volume doubled.