GitHub published a guide teaching coding agents to decompose their output into stacked pull requests — ordered, individually reviewable layers instead of one unreviewable monolith.
The problem is mechanical: AI tools like GitHub Copilot generate high volumes of code fast, and review quality degrades as PR size grows. Reviewers skim, miss issues, or procrastinate until the PR goes stale and develops merge conflicts.
What stacked PRs change:
- Ordered layers — each PR builds on the one beneath, so reviewers see one logical chunk at a time
- Smaller diffs — individual pieces of a feature get reviewed instead of one enormous code change
- Agent-side decomposition — the tutorial teaches agents to produce stacks automatically, not manually split after the fact
NOTE
GitHub’s Copilot tutorial covers both the agent-side decomposition and the review-side workflow for stacked PRs.
AI coding tools — Cursor, Claude Code, GitHub Copilot — generate more code per session than a human can review in one pass. Stacked PRs don’t solve the review capacity problem. They make it tractable by breaking it into pieces.
Available now in GitHub’s engineering blog post and Copilot tutorials.