The rise of the Generative AI The world of software development seemed to follow a clear script: models would write the code and humans would review it. It was the new balance. Well, Anthropic just killed him.
The problem of programming with AI. What we know today as vibe codingthis practice of giving instructions in natural language to an AI so that it generates code at full speed, has skyrocketed software production in companies. Anthropic affirms that the amount of code generated by each of its own engineers has grown by 200% in the last year. And now there’s a problem: there’s so much new code that reviewing it has become the bottleneck of the process.
Human developers can’t cope. Many pull requests (change proposals that must be reviewed before integrating new code) are skimmed or not read very carefully at all.
What Anthropic has done. The company Code Review has been releaseda tool integrated into Claude Code that, instead of waiting for a human to review the code, deploys a team of AI agents to do it automatically every time a pull request is opened. This new system is now available in preview phase for Team and Enterprise plan customers.
Cat Wu, Product Manager at Anthropic, explained told TechCrunch that the question they constantly received from their clients’ technical managers was always the same: “Now that Claude Code is generating a ton of pull requests, how do I make sure they are reviewed efficiently?”
How it works inside. AI agents work in parallel autonomously the moment a pull request is opened, examining the code from different perspectives. An end agent then aggregates and prioritizes the issues it has found, removing duplicates and sorting them by severity. The result reaches the developer through a featured comment, accompanied by more online comments about specific bugs.
The focus, according to Anthropicis in logical errors, not in matters of style, something designed on purpose so that the feedback does not generate too much noise. Issues are labeled by color depending on how important they are: red for critical, yellow for attention, and purple for pre-existing code.
Numbers. The company has been using Code Review internally for months before launching it to the market. According to what they saybefore implementing it, only 16% of their pull requests received meaningful review comments. With the tool, that percentage rises to 54%. In large pull requests (more than 1,000 modified lines) 84% returned results, with an average of 7.5 problems detected. And less than 1% of those results are flagged as incorrect by the engineers themselves.
In one of the cases documented by the company, they spoke of a single line change that seemed routine. However, Code Review marked it as critical, as it apparently could have broken the entire service’s authentication. The bug was fixed before integration. Furthermore, according to the company, the engineer later acknowledged that he would not have caught it alone.
ANDhe new role of the programmer. The narrative that had spread in the last two years was that developers would evolve towards a profile closer to that of a reviewer or supervisor of code generated by AI. Now that transition is also being automated, at least in part. Anthropic does not eliminate the human from the equation (in fact the tool does not approve pull requests), but it does compress the review work that was supposed to be the last bastion. It seems that now the human goes from reviewer to final arbiter.
Price. It is not a cheap tool. Each revision has a cost based on token consumption. Anthropic esteem The average price per review is between $15 and $25, depending on the complexity of the code. It is a cost that the company justifies in the context of large technology companies where errors that escape review have a much higher price.
Cover image | Compagnons
In Xataka | Software companies sank on the stock market for a simple reason: investors are panicking about AI
GIPHY App Key not set. Please check settings