Cognition Introduces "Stacked PRs" in Devin, Splitting Large Changes into Reviewable Layers
Cognition, maker of Devin, has introduced “Stacked PRs,” a feature that automatically breaks large AI-generated changes into a stack of smaller, independently reviewable pull requests instead of one massive PR.
Details
- The problem it addresses: Research cited by Cognition (Faros AI, 2025) found that teams with high AI coding adoption merged 98% more PRs, but review time per PR rose 91% and average PR size grew 154%, with no measurable gain in DORA delivery metrics — AI agents were generating work faster than humans could review it
- How it works: Rather than submitting one monolithic PR, Devin decomposes the change into a “stack” of small, self-contained PRs, each reviewable on its own
- Architectural awareness: The decomposition relies on DeepWiki, Devin’s continuously indexed map of a repository’s modules and the dependencies between them, letting Devin split changes along natural architectural seams rather than arbitrarily
- Automatic rebasing: When a reviewer requests changes on an earlier layer of the stack, Devin automatically rebases the downstream PRs and resolves the resulting conflicts
- Retroactive stacking: Devin can also take an existing large PR and restructure it into an organized stack after the fact
- Stack management: Devin manages CI status, review state, and mergeability across every layer of the stack, not just the individual PR
- Partner: Built together with GitHub
How to try it
- Stacked PRs is available now to Devin users; large tasks submitted to Devin will automatically be organized into a stack when appropriate
- Full details are in Cognition’s blog post at devin.ai/blog/introducing-pr-stacks