AI
Anthropic Shifts Code Review to Risk-Based Model as AI Contributions Surge
Rather than reviewing every line of code, Anthropic's Claude team now gates human approval on risk level, letting automated agents handle routine checks and routine changes.
InfoQ
Anthropic has redesigned its code review process to accommodate the explosive growth in AI-generated code contributions to Claude. The system replaces traditional line-by-line human review with a multi-layered approach combining automated checks, AI-powered analysis, and human approval gates determined by change risk.
The Scale of Change#
The volume of merge requests to the Claude Code repository more than doubled between June and September this year. Despite this surge, the proportion of changes that had to be reverted actually dropped by over 40%—suggesting that the new review infrastructure is effective at catching problems early while letting safe changes through faster.
How the New Process Works#
Before a pull request is even opened, Claude runs a battery of preliminary checks: tests, builds, and linters. Separately, an AI agent reviews the proposed changes in isolation, identifying potential issues upfront.
Once a PR is formally submitted, multiple automated layers kick in:
- Claude-powered reviewers scan for bugs and other defects
- Independent verification agents double-check findings
- Security reviews can block merges if unresolved concerns remain
- Review cycles are capped at three rounds to prevent endless feedback loops
Rather than examining the full changeset repeatedly, later review rounds focus only on new modifications made by the author.
Risk-Based Human Oversight#
The key innovation is calibrating human involvement to change risk:
- Low-risk changes (documentation, feature-flagged internal work) may skip human review entirely
- Medium-risk changes require one human approval
- High-risk changes need two approvals
- Sensitive areas (CI workflows, core systems) always get human review
This shift has measurable benefits. The share of PRs passing on the first review-and-fix cycle rose from 65% to 78%, while the proportion requiring five or more rounds fell from 8.9% to 1.5%.

