Plan
The ticket becomes a plan: work steps with dependencies, open points as visible assumptions. You see what will be built before it is.
The principle
An AI software factory takes a ticket in at one end and delivers a draft pull request at the other, in which every work step was produced under the same rules, built and tested, and leaves its own commit. Your team stands at the end of the line and decides. In between, the factory works.
Six agent PRs of 800 lines each in the queue, no trace of why the agent built it that way, stand-up in an hour. The tools that were supposed to make things faster only moved the work: from typing into review. More code, the same judgement, the same person.
A factory does not solve that with a bigger prompt but by cutting to size. Every ticket becomes a chain of small work steps, each with exactly the context it needs. A coordinator holds the plan, workers build the steps in parallel, and barriers between the phases make sure nothing moves on that is not finished.
The ticket becomes a plan: work steps with dependencies, open points as visible assumptions. You see what will be built before it is.
Every step runs on its own worker with its own worktree, under the same guardrails as all the others, and leaves its commit.
Build and tests run in a sandbox with your project’s commands. What fails is repaired. Where the toolchain is missing, the run says so instead of claiming green.
The result is not a chat transcript somebody copies code out of. It is a draft pull request in your repository with a history: one commit per step, every check recorded. You read what was asked, what changed and what was checked, and merge if it fits.
What a good first assignment looks like, and what you want to see before you merge.
Follow one assignment from the tracker to the draft pull request in your repository.
Turn a requirements text into epics and stories your team can prioritise.