A fast workflow that fails quietly is worse than a slow one that tells you what went wrong. So every new step type we ship starts with one question: what happens when this breaks at three in the morning?
Failure is a feature
Each step declares how it can fail, how often it may retry and who hears about it. Those rules are visible on the canvas, not buried in settings, so a reviewer can read them like any other part of the workflow.
Only once a step fails well do we make it fast. Most of the time the two goals agree, and when they do not, we pick the boring, predictable option.
