Methodology

I architect full software development cycles with AI, keeping them under control through Gubia, a harness-agnostic framework I built on top of a simple idea, the Ralph Loop.

Ralph’s approach is to repeat a task until the model eventually completes it correctly. That works for simple, measurable problems. Add complexity or a decade-old legacy codebase and the cracks show: errors, hallucinations and a very large token bill. I kept that core idea and built on it so development can continue through a whole cycle with far fewer errors.

Gubia starts with a planning phase that produces a plan a human can audit, breaking the cycle down into tasks and subtasks for the implementing agents. Inside the plan, a self-review framework of LLM judges and evidence means a subtask only counts as done when it can be shown. The plan also documents itself: discoveries made along the way are recorded, so the models don’t repeat past mistakes or drift from the path already taken.

It has worked well on projects from scratch to a ten-year-old legacy codebase. Even so, I recommend reviewing the result as you would with any PR.

If you want to know more, take a look at Gubia.