Summary
Anthropic recently published a post describing how the team behind Claude Code reviews its own code, a notable challenge given that Claude writes roughly 90% of the code in Claude Code. This video from David Shapiro walks through that process and what it suggests about the future of software teams.
The breakdown covers the numbers Anthropic shared: pull requests merged into Claude Code more than doubled between June and September, while the share of changes that had to be reverted fell by more than 40%. Merged changes in Anthropic’s monorepo grew from about 2,500 a day in May to more than 30,000 by September. The review pipeline layers verification, multi-agent AI review with voting, automated security review, standard tests and builds, and risk-tiered human approval. The share of pull requests receiving substantive review comments rose from 16% to 54%, and single-round pass rates climbed from 65% to 78%.
Shapiro adds a note of caution, since Anthropic is evaluating its own results and the figures haven’t been independently verified. He then argues the underlying idea is far older than AI, pointing to database logs, TCP checksums, error-correcting memory, double-entry bookkeeping, and Claude Shannon’s information theory. His conclusion: reliable systems come from machines checking machines, with humans designing the checks and remaining accountable for outcomes.
📺 Source: David Shapiro · Published October 24, 2026
🏷️ Format: News Analysis
