Turn one giant AI-generated pull request to a reviewable stack

GitHub reports that developers often face a dilemma when using coding agents: either accept one massive pull request that is difficult to review or manually split work into smaller, interdependent requests prone to conflicts. The company argues that neither approach scales well as AI-driven productivity gains accelerate. According to Gartner, coding agents are projected to boost software development lifecycle productivity by 50% by 2028, intensifying the need for better review structures. GitHub proposes stacked pull requests as a solution, enabling agents to decompose large changes into logical, independently reviewable layers while maintaining dependency order.

The workflow begins by setting a stack base, which serves as the reference point for CI checks and merge rules throughout the process. Developers then identify the foundational unit of work and place it at the lowest layer, with dependent changes stacked above. This decomposition clarifies ownership—data changes reviewed by data specialists, UI changes by design teams—and reduces cognitive load for reviewers. GitHub’s native support for stacked pull requests is accessible via the pull request UI and the gh stack CLI, integrating with existing workflows without requiring new tools.

To automate this process, GitHub introduces the gh-stack skills, which teach coding agents how to create and manage stacked pull requests. For the shopping assistant feature example, custom agents follow strict scoping rules to generate small, single-purpose pull requests. This ensures each layer remains focused and reviewable, preventing the sprawling diffs that typically overwhelm reviewers. The approach aligns with how teams traditionally structure code but adapts it for AI-assisted development.

Each pull request in the stack is evaluated against the stack base, with CI checks running independently for every layer. GitHub emphasizes that while many agent workflows operate autonomously, the stacked structure allows for step-by-step verification, reducing the risk of conflicts and overlooked changes. The company positions this as a practical way to integrate AI-generated code into collaborative workflows while preserving review quality and maintainability.