Classical Machine Learning models have problems when faced with learning from data that changes over time or across domains due to factors such as noise, occlusion, illumination, or frequency variation, which humans can adapt to without being given independent and identically distributed data. Therefore, a Continual Learning (CL) approach is essential, particularly in this case, the Domain-Incremental Learning. This paper presents a novel pipeline for identifying tasks in domain-incremental learning scenarios without supervision. The pipeline consists of four steps. First, we obtain base embeddings from the raw data through a transformer-based existing model. Second, we group the embedding densities based on similarity and extract few nearest points to each cluster centroid. Third, we train an incremental task classifier using only the previous few points. Finally, we make the pipeline lightweight in terms of computational requirements and build an algorithm that decides when to learn a new task in an online way using the task classifier and drift detector. We experiment with the real-world driving dataset SODA10M and several CL strategies. We conclude that the performance of CL strategies with our pipeline is better not only when the task boundaries are given, but also in the more general practical case of task-agnostic strategies that demand identifying new tasks on-the-fly.