The Inspector's Instinct

A critique of how teams govern AI agent systems by replicating human-worker oversight patterns (approval gates, kill switches, escalation ladders) without understanding whether those controls address real failure modes. Drawing on Deming's System of Profound Knowledge, the author argues that most governance scaffolding manages organizational anxiety rather than actual risk. The alternative proposed is observation-first design: using a typed, versioned data layer (illustrated via the Swamp framework) to make agent behavior, variation, and decision context queryable and traceable. Controls should be derived from observed failure modes, not guessed in advance. The post also acknowledges where registry-and-manifest approaches add genuine value (cross-domain discoverability) while arguing that declared intentions are a poor substitute for observed capabilities.