Dynamically adapting provision of notification output to reduce distractions and/or to mitigate usage of computational resources. In some implementations, an automated assistant application predicts a level of engagement for a user and determines, based on the predicted level of engagement (and optionally future predicted level(s) of engagement), provisioning (e.g., whether, when, and/or how) of output that is based on a received notification. For example, the automated assistant application can, based on predicted level(s) of engagement, determine whether to provide any output based on a received notification, determine whether to suppress provision of output that is based on the received notification (e.g., until a later time with a decreased predicted level of engagement), determine whether to provide output that is a condensed version of the received notification, determine whether to automatically respond to the notification, and/or select an output modality for providing output that is based on the received notification.