If you've ever taken code from an agent, you know the picture: the task is one thing, and the diff is two thousand lines across eight files. Technically it's all there. Practically there's nowhere to grab hold of it.
Fact
The facts: on 4 August GitHub published an engineering write-up and a tool - gh-stack, an extension for the gh command line (gh extension install github/gh-stack). The idea: instead of one big pull request, the work is arranged into a stack of small, logically sequential, independently reviewable layers. Commands: init, add, push, submit, rebase, sync. GitHub's example: product search in a shopping assistant, which as a single PR ran over 1720 lines, splits into four layers - the data catalog, the search API, the chat connection, the interface. Review runs top-down for context and bottom-up for approval; a change in a middle layer cascades to the dependent ones. GitHub cites a forecast of a 50% productivity gain from agents across the whole development cycle by 2028. Source: github.blog, 04.08.2026.
This is where the craft shows, and the admission along with it. When GitHub builds a tool for slicing up AI diffs, that's an official admission of where the bottleneck has moved - no longer writing, but reading. The only trust pipeline the craft has carried for decades is one person reviewing another's work - piece by piece, at a size the brain can carry.
The good thing about the design: the stack stands on ordinary branch-based work and ordinary pull requests, request-by-request. Nothing new to learn. The same git, arranged with discipline. Tools like that survive.
Code got cheaper. Review didn't.
In our AI workflows you plan the layers before the agent sits down to write, not after it hands you the pile. You cut it into steps right in the brief - then the stack comes out on its own, and review stops being a punishment.
Official primary sources
GitHub Engineering - Turn one giant AI-generated pull request into a reviewable stack, 04.08.2026The visual is generated code-art, no third-party images.