Pizzaro is my computer-use harness. It turns a task prompt into UI actions, and then earns the right to replay them without a model.
A planner breaks a task prompt into an ordered plan of clicks, typing, scrolls and keys, reusing actions it already knows. A human confirms or edits each step before it becomes part of a knowledge graph of actions.
Actions move from Exploring to Verifying to Trusted as they pass verified replays. Trusted actions replay programmatically using template matching and cheap pre/postchecks that need no model. When one fails during an autonomous run, it is demoted to a repair backlog with the screenshot at failure, so it can be fixed and re-baked.