Cognition has introduced Code Scans, a Devin feature intended to investigate broad engineering goals across an entire repository and convert selected findings into pull requests. Rather than requiring a developer to identify a particular file or defect first, the tool starts with an outcome such as finding unused code, checking accessibility rules or preparing for a migration.
Users define the goal and refine the boundaries with Devin, including which files to inspect, what to exclude and what should qualify as a finding. Code Scans then divides the investigation into focused batches, sends those batches to parallel agents and combines their results in a prioritized report. Developers can review the evidence, discuss individual findings and ask Devin to implement chosen changes as pull requests.
The architecture builds on what Cognition calls Agentic MapReduce, previously used for its Security Swarm product. The approach is meant to keep each agent focused on a limited portion of a large repository while a coordinating stage creates a unified result. That structure addresses a common obstacle with open-ended maintenance work: identifying all relevant locations can itself require a substantial engineering project before any fix begins.
Cognition reported a 96% pull-request merge rate across several repositories during pre-release testing and estimated that the feature saved more than 700 engineering hours. Those are company estimates, not independently audited measurements, and the release did not provide a standardized comparison against human review or other tools. They nevertheless indicate how Cognition expects customers to use the feature: for repeated, repository-wide maintenance rather than only isolated code generation.
One test targeted compilation time in the Dioxus repository. Cognition says Code Scans identified heavy default features and dependency configuration that could be made optional. After proposed changes were applied locally, a clean debug build across 22 workspace crates fell from 58.6 seconds to 21 seconds, a reported 64% reduction. The company also ran scans for search-engine optimization issues across its Devin and Cognition websites, reporting 44 findings involving matters such as metadata, redirects and structured data.
The examples show that the quality of the initial criteria still matters. A request to remove unused code, for instance, can specify that generated files and test fixtures be excluded and that indirect references be checked before deletion is recommended. Developers remain responsible for evaluating evidence and deciding which changes should be merged.
Code Scans is available through the Devin web application using the `/scan` command. The launch extends the product from handling defined implementation tasks toward examining a repository for opportunities that have not yet been converted into tickets, while retaining pull-request review as the decision point before changes enter a codebase.
