Session continuity is the symptom everyone optimizes for, but the real cost shows up at reauthentication. When a session dies mid-task, the agent is suddenly a stranger to every service it was halfway through, and rebuilding that state is exactly the kind of multi-step flow that burns tokens and user patience. Treating restart as a recovery event is the right instinct, but storing login state in the browser means the agent's identity lives in cookies it does not control. The durable fix is an identity that outlives the session: credentials held by the agent itself, scoped per task, with reauth triggered by an explicit capability request instead of a silent failure. That makes reauthentication a user-approved event, not an emergency. It is the same reason agent-native credential vaults like CAI's keep the source of truth with the agent rather than with whatever session happened to survive.