Recently, deep learning methods have been successfully applied to solve numerous challenges in the field of digital pathology. However, many of these approaches are fully supervised and require annotated images. Annotating a histology image is a time-consuming and tedious process for even a highly skilled pathologist, and, as such, most histology datasets lack region-of-interest annotations and are weakly labeled. In this paper, we introduce HistoPerm, a view generation approach designed for improving the performance of representation learning techniques on histology images in weakly supervised settings. In HistoPerm, we permute augmented views of patches generated from whole-slide histology images to improve classification accuracy. These permuted views belong to the same original slide-level class but are produced from distinct patch instances. We tested adding HistoPerm to BYOL and SimCLR, two prominent representation learning methods, on two public histology datasets for Celiac disease and Renal Cell Carcinoma. For both datasets, we found improved performance in terms of accuracy, F1-score, and AUC compared to the standard BYOL and SimCLR approaches. Particularly, in a linear evaluation configuration, HistoPerm increases classification accuracy on the Celiac disease dataset by 8% for BYOL and 3% for SimCLR. Similarly, with HistoPerm, classification accuracy increases by 2% for BYOL and 0.25% for SimCLR on the Renal Cell Carcinoma dataset. The proposed permutation-based view generation approach can be adopted in common representation learning frameworks to capture histopathology features in weakly supervised settings and can lead to whole-slide classification outcomes that are close to, or even better than, fully supervised methods.