Alibaba Cloud introduces Kitta, a domain-focused code review agent

A response to growing review pressure

Alibaba Cloud presents Kitta against a shift in software development: AI-assisted programming is increasing code output, while review capacity has not grown at the same pace. The company says the bottleneck in research and development is moving from writing code to reviewing it.

The article cites DORA's annual study, State of AI-assisted Software Development, which surveyed about 5,000 respondents. According to the figures cited by Alibaba Cloud, deeper AI involvement in development correlates with a 9% rise in code bug rate, a 91% increase in code review time and a 154% increase in average pull request size. Alibaba Cloud also says these figures are correlation measures rather than causal conclusions.

The source describes code review as a central stage in OpenAnolis' planned Agentic Continuous Integration pipeline. It sits between AI-assisted programming and later intelligent testing and community merging, with people retaining the final merge decision.

Why general-purpose review tools fall short

Alibaba Cloud says general-purpose AI review tools can respond quickly and intercept basic defects, but often lack the project understanding required for expert-level work. The company identifies problems including incomplete coverage, inaccurate locations for reported issues and unstable results when prompts change slightly.

The deeper issue, according to the article, is that different projects require different technical and business judgments. A kernel review may depend on locking semantics, while a compiler review may require knowledge of conventions around its intermediate representation. Standards can also differ between an open-source community project, an internal enterprise repository and distribution maintenance.

Kitta addresses this by turning a general-purpose agent into a domain specialist. Its process begins with real review and defect records, including data from project repositories. These records are collected, associated, normalized, denoised, semantically parsed and independently reproduced before becoming project-level expert assets.

Reported results across six domains

The resulting assets include judgment criteria, such as review principles, defect patterns and project conventions, along with validation samples covering problematic code, normal code and expected conclusions. Alibaba Cloud says this approach makes domain knowledge verifiable, reusable and capable of being developed into an organizational capability rather than remaining personal experience.

Kitta has been running in the OpenAnolis community since it went online. Alibaba Cloud reports that it has completed reviews of more than 70,000 patches, with a gatekeeping accuracy of 97% and a merge rate stable at 62%.

The company says its domain-customization methodology has been validated in six domains, including operating-system kernels, compilers and Agent platform applications. It reports improvements in review quality ranging from 17.0% to 31.6%, while freeing maintainers from line-by-line review so they can focus on disputes that require human judgment.

For users, the announcement points to a model in which AI review is customized around a project's history, standards and validation data. Alibaba Cloud's figures are company-reported, but they indicate the intended role for Kitta: handling more of the routine review burden while keeping community maintainers responsible for final decisions.