Cognition has introduced Code Scans, a new feature for its AI engineer, Devin, designed to transform broad engineering goals into concrete codebase improvements. The tool allows users to specify high-level objectives—such as finding unused code or identifying SEO issues—and automatically investigates the codebase to find actionable items.
Product LaunchesCognitionDevinCode Scans
Devin Introduces Code Scans for Automated Codebase Investigation and PR Generation
Code Scans is built on Agentic MapReduce, an architecture originally developed for Devin Security Swarm. This architecture breaks large investigations into focused batches and distributes them across parallel agents, synthesizing the findings into a single report. Once the investigation is complete, Devin can transform these findings into pull requests (PRs) ready for review.
In early testing, the tool has demonstrated significant practical impact. For example, when used to speed up compilation in the Dioxus repository, Devin identified dependency and build configuration optimizations that reduced clean debug build times from 58.6 seconds to 21.0 seconds, a 64% reduction. Additionally, testing showed an estimated 96% PR merge rate across several repositories, saving over 700 engineering hours during the testing period.
Sources
- Code Scans （Hacker News Frontpage, 2026-09-18）
- docs