← Back to feed

Cognition has added Code Scans to Devin, targeting the class of engineering work that never gets started because the first step — figuring out where to begin — is itself a substantial project.

Code Scans accepts a goal rather than a file list. Tell Devin “improve SEO,” “reduce maintenance overhead,” or “make compilation faster,” and it investigates the entire codebase, produces a findings report, and converts those findings into pull requests ready for review.

Architecture: Agentic MapReduce

The underlying mechanism is Agentic MapReduce, the same architecture Cognition built for Devin Security Swarm. Devin breaks a large codebase investigation into focused batches, distributes them across parallel agents, and synthesises the results into a single report. The fan-out handles the discovery phase; the synthesis step produces actionable output rather than raw findings.

The feature is triggered via /scan in the Devin webapp.

Early Numbers

Philips Digital Computational Pathology, Enterprise Informatics, ran Code Scans across several repositories before launch. Results:

• 96% PR merge rate across scanned repos
• 700+ engineering hours saved during the testing window

Cognition describes the metric as covering “software quality, bug fixes, stability, performance, and compliance including security and privacy” — not just cosmetic changes.

A 96% merge rate is a meaningful bar. Automated PR generators typically produce high-volume, low-acceptance output. The constraint here is that Devin runs a scoped investigation first, then generates targeted changes, rather than pattern-matching across files without context.

What It Changes

The prior workflow required an engineer to (1) scope the problem, (2) identify affected files, (3) implement changes one-by-one. Code Scans collapses steps 1 and 2 into a parallel agent sweep, then feeds the map phase results into a reduction layer that generates concrete changes.

For backlogs of outcome-defined work — compliance posture, dependency hygiene, dead code removal, accessibility conformance — Code Scans offers a path to implementation without requiring an engineer to own the investigation phase.

Devin’s Commercial Context

Cognition raised $1B at a $26B valuation in May 2026, with Devin ARR reported at $492M at the time. Code Scans extends the same agentic engine that drove that growth into a new task class: large-scope codebase work where the goal is defined at the outcome level rather than the task level.