The analysis of task-related fMRI data at the level of individual participants is commonly based on general linear modeling (GLM), which allows us to estimate the extent to which the BOLD signal can be explained by the task response predictors specified in the event model. The predictors are constructed by convolving the hypothesized time course of neural activity with an assumed hemodynamic response function (HRF). However, our assumptions about the components of brain activity, including their onset and duration, may be incorrect. Their timing may also differ across brain regions or from person to person, leading to inappropriate or suboptimal models, poor fit of the model to actual data, and invalid estimates of brain activity. Here, we present an approach that uses theoretically driven models of task response to define constraints on which the final model is computationally derived using actual fMRI data. Specifically, we developed autohrf–an R package that enables the evaluation and data-driven estimation of event models for GLM analysis. The highlight of the package is the automated parameter search that uses genetic algorithms to find the onset and duration of task predictors that result in the highest fitness of GLM based on the fMRI signal under predefined constraints. We evaluated the usefulness of the autohrf package on two original datasets of task-related fMRI activity, a slow event-related spatial working memory study and a mixed state-item study using the flanker task, and on a simulated slow event-related working memory data. Our results suggest that autohrf can be used to efficiently construct and evaluate better task-related brain activity models to gain a deeper understanding of BOLD task response and improve the validity of model estimates. Our study also highlights the sensitivity of fMRI analysis with GLM to precise event model specification and the need for model evaluation, especially in complex and overlapping event designs.