Anthropic launches a code review tool to verify the flow of AI-generated code

When it comes to programming, peer feedback is crucial for catching bugs early, maintaining consistency across the code base, and improving overall software quality.

The emergence of “dynamic coding” — using AI tools that take instructions presented in simple language and quickly generate large amounts of code — has changed the way developers work. Although these tools sped up the development process, they also introduced new bugs, security risks, and incomprehensible code.

Anthropic’s solution is an AI reviewer designed to detect errors before they reach the software database. The new product, called Code Review, was launched on Monday in Claude Code.

“We’ve seen a lot of growth in Cloudcode, especially within the enterprise, and one of the questions we’re constantly getting from enterprise leaders is: Now that Cloudcode has put out a bunch of pull requests, how can I make sure those requests are reviewed in an efficient way?” Kat Wu, head of product at Anthropic, told TechCrunch.

Pull requests are a mechanism used by developers to submit code changes for review before those changes enter the software. Wu said Claude Code It led to a significant increase in code production, which led to an increase in pull request revisions that caused shipping code to bottleneck.

“Code review is our answer to that,” Wu said.

Anthropic’s launch of Code Review — which arrives first to Claude for Teams and Claude for Enterprise customers in Search Preview — comes at a pivotal moment for the company.

Monday Anthropy Two lawsuits were filed v. Department of Defense in response to the agency’s designation of Anthropic as a supply chain risk. The dispute is likely to leave Anthropic relying more heavily on its booming business, which has seen subscriptions quadruple since the start of the year. Claude CodeRun rate revenue has exceeded $2.5 billion since launch, according to the company.

“This product is very much aimed at our large-scale enterprise users, so companies like Uber, Salesforce and Accenture, who are already using Claude Code and now want to help with the huge amount of (pull requests) that it helps produce,” Wu said.

Developer leaders can turn code review on by default for every engineer on the team, she added. Once enabled, it integrates with GitHub and automatically analyzes pull requests, leaving comments directly on the code pointing out potential issues and suggested fixes.

Wu said the focus is on fixing logical errors rather than style.

“This is really important because a lot of developers have seen automated AI feedback before, and they get upset when it’s not immediately actionable,” Wu said. “We decided that we would only focus on logical bugs. That way we would pick up the highest priority things to fix.”

The AI ​​explains its causes step-by-step, explaining what the problem is, why it might be a problem, and how it can be fixed. The system will classify the severity of issues using colors: red for highest severity, yellow for potential issues worth reviewing, and purple for issues related to pre-existing code or historical bugs.

It does this quickly and efficiently by relying on multiple agents working in parallel, with each agent examining the code base from a different perspective or dimension, Wu said. The final agent aggregates and sorts the results, removing duplicates and prioritizing what is most important.

The tool provides light Security analysisand engineering leads can customize additional checks based on internal best practices. Wu said Anthropic recently launched Cloud security code Provides deeper security analysis.

The multi-agent architecture means this can be a resource-intensive product, Wu said. Similar to other AI services, pricing is based on codes, and the cost varies depending on the complexity of the code — though Wu estimated that each review would cost $15 to $25 on average. She added that it is a distinct and necessary experience because artificial intelligence tools generate more and more code.

“(Code review) is something that comes from a tremendous amount of market traction,” Wu said. “As engineers evolve with CloudCode, they see the difficulty of building a new feature (attrition), and they see a much higher demand for code review. So we hope that through this, we will enable organizations to build faster than ever before, and with far fewer errors than they did before.”