We do a fixed number of retries, including redoing any AI arguments. We've thought about making it atomic/more durable -- it's tricky, given that most steps interact with external systems e.g. Google Sheets, and while not typically "destructive" (Google Sheets has version history), undo-ing is often difficult.