This study delves into the active aspect of the sense of agency (SoA), often underrepresented in literature, positing it as an active causal inference outcome concerning one's actions and their environmental impact. Participants engaged in computer-mouse-driven tasks, designed to evaluate their control or recognition of controlled visual objects amidst varying noise levels. Our findings indicate that participants actively crafted high-level, low-dimensional action plans, unique yet consistent within individuals, to deduce their control level in a noisy environment. Utilizing transformer-LSTM-based autoencoders, we quantified these action plans, revealing that their geometrical and dynamical attributes could significantly predict task-related behavioral profiles. This implies that participants' sense of control is molded by actively modifying action plans, which are seen as tools for generating causal evidence via intervention. Additionally, participants actively broadened their action plan diversity, as indicated by the action plan distribution's dimensionality. This approach promotes exploration of potential action plans while simultaneously gathering causal evidence for inference. Contrastingly, schizophrenia patients showed lesser action plan diversity, hinting at diminished active control inference and weaker self-relevant cue detection. However, their judgment responses were in line with predictions based on the geometrical and dynamical features of their action plans, suggesting they adjust their decision-making boundaries to enhance their active causal inference regarding environmental changes. These insights contribute to a richer understanding of SoA, underscoring its roots in active causal inference.