The bad versus sad split is the part I keep thinking about. In integration work like Kombo's, almost everything an agent gets subtly wrong shows up as a sad first: one customer's field mapping is off, nobody files a ticket, and it only becomes a bad three setups later. Writing the quality bar into the repo helps, but the harder gap is that sads rarely get reported at all, so the model never sees the standard it just missed.
The bad versus sad split is the part I keep thinking about. In integration work like Kombo's, almost everything an agent gets subtly wrong shows up as a sad first: one customer's field mapping is off, nobody files a ticket, and it only becomes a bad three setups later. Writing the quality bar into the repo helps, but the harder gap is that sads rarely get reported at all, so the model never sees the standard it just missed.