AI writing code isn’t the bottleneck. Review is.

WHAT HAPPENED

Anthropic explained how it does code review in Claude Code

The team describes a layered flow: checks before the PR, review by agents, security review, mandatory CI, human approval based on risk, merge queue, flags, and fast rollback. The idea is to save human time without increasing reverts or incidents.

THE NUMBER

40%

more than a 40% drop in the share of reverted PRs

From June to September, while merged PRs more than doubled.

THE REAL PROBLEM IN COMPANIES

What I see in companies when they try to speed up with AI

1. 01AI generates more code, but review stays manual and turns into a queue
2. 02The team automates writing, but doesn’t integrate validation into the real workflow
3. 03Without simple rollback, every speed gain turns into production risk

I’VE ALREADY SOLVED THIS

This is how we did it on the AskMoses.AI development team

1. 01each dev works with Claude Code on the project
2. 02another Claude reviews the pull request
3. 03deploy on Vercel with 1-click rollback and daily backup

MY TAKE

Measure instead of assuming.

If AI speeds up writing but increases rework, incidents, or approval wait time, there was no real gain. In this case, the point isn’t how many agents write code. It’s how much of the flow becomes reliable, auditable, and reversible in production.