The white paper favors using access tokens for agents that are extremely narrow in scope, extremely short in validity, and bound to specific task contexts.