The software factory described here is not a particular model, coding harness or commercial product. It is a model-agnostic operating system for AI-assisted development, encoded in project instructions and reusable skills so agents follow the same development process each time.
The workflow begins by isolating every feature in its own Git worktree and branch, which lets multiple agents work in parallel without overwriting one another. A code-structure skill then steers implementation toward a maintainable service-layer architecture that people and later agents can understand.
Verification is treated as part of the build rather than a final promise. Agents capture before-and-after evidence or measured test results, then an external code-review agent scores the pull request and sends unresolved issues back through the build and proof stages until the required quality threshold is met. Human intervention is reserved for the final review and merge.
Watch on YouTube
