Claude writes ~90% of Claude Code's own code. So how does Anthropic review it? Not by having a human read every line.
Their team published their setup, and one idea stuck with me: "did a human read it?" is the wrong safety check. The real question is what could break, and which check should catch it.
Automated checks first
Every PR goes through automated checks before a human sees it. Claude has to prove its own work with tests and screenshots. Several AI reviewers hunt for bugs and vote, and only the agreed findings get posted. Security issues block the merge.
Humans scaled to risk
- Low risk (docs, copy, flagged additive changes): zero human approvals- Medium (normal product changes): one approval- High (billing, auto-updater, CI workflows): two approvalsA classifier sets the tier from the change, not the author. And an agent's approval never counts as a person's. ## Results Merged PRs doubled in three months. Reverts dropped 40%. PRs passing review in one round went from 65% to 78%. ## The bit I'm stealing Every past incident becomes an eval for the AI reviewer. If a prompt change doesn't catch more old bugs, it doesn't ship. I run something similar with my own agents, and the humans don't go away. They just stop reading lint output and start making the calls machines can't. Source: the Claude Code team's post on claude.dev, October 2026.
Top comments (2)
The reframing from "did a human read it?" to "what could break, and which check should catch it?" is a really useful one. Turning every past incident into an eval for the AI reviewer is the part I'd copy too, since it gives the reviewer a memory of your own team's failure modes. Did you find the risk-tier classifier needed much tuning at first, or did simple path-based rules get you most of the way?
The 90% matters less than what the reviewer is allowed to stop. A check that can only leave a comment gets skipped once the queue grows. A check that can fail the merge does not.