The natural outcome of the code generation boom: engineers are cranking out code much faster (Anthropic directly states a 200% increase in volume per person per year), and there's physically no one to review this mass.
Anthropic saw this and rolled out Claude Code Review (currently in preview for Team/Enterprise). It's a full-fledged multi-agent system.
How it works: a swarm of agents is unleashed on each open PR. They search for vulnerabilities and logical errors in parallel, then a verification phase filters out false positives (by cross-checking with actual code behavior), and finally inline comments are generated. On average, such a deep run takes about 20 minutes.
Average cost per review: $15–$25 (deducted in tokens). 😱
There are trigger settings. If you habitually enable "check on every push" to run agents on every typo-fix commit, you'll soon have to sell all your kidneys.
To customize the nitpicking, you can place a REVIEW.md file in the repository root. There you write the team's rules in human language.
The neural network also highlights dormant bugs in neighboring code that the PR merely touches tangentially.
According to Anthropic's internal statistics, in large PRs (>1000 lines), agents find critical issues in 84% of cases. Meanwhile, the code authors themselves reject less than 1% of the comments.
Could you have imagined a couple of years ago that enterprise development would mainly work like this: neural networks generate code by the ton, other neural networks (at a very high cost) review that code and find bugs.
And the human meatbag's role in this scheme is just to pay the bills 🥲
Comments
0No comments yet.