Anthropic’s Claude Code team has fundamentally restructured its engineering workflow, moving away from traditional line-by-line human review to a multi-layered automated system. Between June and September 2026, merged pull requests to the Claude Code codebase more than doubled, yet the share of changes requiring reverts fell by over 40 percent. This data point challenges the prevailing assumption that increased AI-generated code volume inevitably leads to lower quality or higher instability.

From Gatekeeper to Continuous Spellcheck

The team redefined the purpose of code review, treating it less as a final gate and more as a continuous process similar to spellchecking. In Anthropic’s broader monorepo, daily changes surged from approximately 2,500 in May to more than 30,000 by September, a scale where manual review becomes impossible. The new workflow ensures most changes are reviewed and fixed by Claude agents before a human is ever asked to look, preserving scarce human attention for high-level architectural questions and novel challenges rather than routine syntax errors.

Six Layers of Automated Defense

Changes now pass through six distinct layers, starting with verification before a PR exists. Engineers provide Claude with runnable tests or design screenshots, and subagents review diffs with fresh context using internal skills for bug hunting and simplification. Boris Cherny, the creator of Claude Code, emphasizes that giving agents a way to verify their own work improves results by two to three times. Every pull request then undergoes AI review against the entire repository, with findings verified by multiple agents voting on their validity to reduce noise.

Risk-Based Human Approval and Convergence

Human intervention is now strictly tiered by risk, with low-risk changes requiring no human approval once automated checks pass, while high-risk areas like billing or CI workflows demand two people. To prevent agent-on-agent loops, the team implemented a round cap; after three review cycles, the reviewer asks the author to reconsider the approach. This convergence strategy was necessary after git blame revealed that 59 percent of flagged issues were pre-existing code, indicating the reviewer was recycling old bugs rather than finding new ones.

Key Takeaways

  • AI review comments on internal PRs jumped from 16% to 54% after the research preview launch, proving agents can catch substantive issues.
  • Security reviews now block merges until findings are resolved, integrating automated safety checks directly into the CI pipeline.
  • Incidents are converted into daily evaluation sets, allowing the team to score reviewer changes against past failures to improve recall.
  • Human roles have shifted to defining direction, taste, and scope, with agents handling the mechanical execution and initial verification.

The Bottom Line

Anthropic isn't just automating code review; they are solving the scalability crisis of AI-generated code by making human attention the bottleneck for decision-making rather than inspection. The drop in reverts suggests that when verification is built into the generation loop, speed and stability are not mutually exclusive.