Devin Scans the Whole Repo First

Most coding agents want a file. Devin's new move wants an outcome.

On 16 September 2026 Cognition shipped Code Scans. You type /scan in the Devin web app and say what you actually care about: unused code, faster compiles, fewer N+1 queries, missing alt text. You do not have to know which path to open. That is the product. The investigation was the part that sat on the backlog.

Image: Cognition, Introducing Code Scans, 16 September 2026.

MapReduce, with agents

The architecture is the same one Cognition used for Devin Security Swarm. They call it Agentic MapReduce, which is a mouthful and also an accurate description.

Plan: Devin reads the repo and writes rules for what counts. Shard: those rules split the matching code into batches. Map: parallel Devins inspect each batch and can read around it. Reduce: one agent merges the notes, drops duplicates, and ranks a report. Then you pick what becomes a pull request.

Completeness is the feature they are selling. Every selected batch has to run. A single agent wandering a monorepo with a huge context window can skip a folder and still sound confident. A map step that must finish the queue is a different failure mode: slower, more complete, more billable.

Two numbers, both from the vendor

Cognition published an in-house compile scan on Dioxus. Devin flagged default features dragging web-sys onto every renderer, an oversized HTML crate, and cargo-generate compiled for a command most people never run. After they applied the changes locally, a clean debug build on 22 workspace crates went from 58.6 seconds to 21.0 seconds. That is a 64 percent drop on one repo they chose. Reproduce it on yours before you quote it.

They also scanned their own sites for SEO issues, found 44 items, and say Ahrefs health on devin.ai moved from 87 to 92 after the fixes. Slow pages down 73 percent. Missing alt text gone on cognition.com. Again: their properties, their crawl.

Philips' Digital Computational Pathology group, testing before launch, is quoted at a 96 percent PR merge rate and more than 700 engineering hours saved. That is a customer quote in a launch post. Keep it in that box.

When this is the right tool

Use a scan when the question is "where?" not "how do I write this function?" Dead code, coverage holes, accessibility against a checklist you provide, a migration map, query patterns. The human job is the scope: exclude generated files, demand evidence of indirect references, say what a finding is allowed to be.

The human job after the report is still merge judgment. A 96 percent merge rate can mean the diffs were small and obvious. It can also mean nobody is reading them. Code Scans does not remove that choice. It just makes the pile.

Sources

Cognition, Introducing Code Scans, 16 September 2026. https://devin.ai/blog/introducing-code-scans