Cognition Code Scans Records 96% Merge Rate and 700 Engineering Hours Saved

Cognition released Code Scans on September 16 2025. The feature converts outcome-level goals such as compile-time reduction or unused-code removal into parallel investigations. It reuses the Agentic MapReduce scheduler previously deployed for Devin Security Swarm, partitioning scans across agents then synthesizing ranked findings into reviewable PRs.

Test data shows a 64% reduction in Dioxus clean debug build time from 58.6 s to 21.0 s after dependency and feature-flag changes across 22 crates. Aggregate metrics report 96% PR acceptance and more than 700 hours saved. These figures exceed typical single-file LLM edit acceptance rates reported in GitHub Copilot studies and align with batch-processing patterns in the 2024 Agentic MapReduce arXiv preprint.

Coverage of the announcement omitted scaling limits on context-window size and the absence of formal soundness guarantees for generated refactors. Parallel agent coordination reduces wall-clock time but inherits verification debt when findings rely on static heuristics rather than exhaustive analysis. Operational impact appears in backlog triage rather than day-to-day editing loops.

Next deployments will likely target compliance and migration scans where rule sets can be supplied by users. Integration depth with existing CI systems remains undocumented.