Cogent: A Modular Cognitive Architecture for Agentic User Simulation

User simulation can support interface design, accessibility research, and interactive agent evaluation, but standard computer-use agents are optimized for task comple tion rather than reproducing how people perceive, reason, and act. We introduce COGENT (COGnitive agENT), a modular architecture for simulating user interac tion with desktop interfaces. Drawing on established cognitive theories, such as Norman’s action cycle and ACT-R, the architecture organizes user simulation into modular processes with explicit representations of perception, memory, user state, intentions, and action selection. We evaluate behavioral fidelity using AgentNet Bench and A11y-CUA, comparing our cognitive agent with direct computer-use agents with and without user-simulation prompting, and evaluate attribute fidelity using SENSE-42. Our agent achieves comparable or higher human-behavior repli cation than the prompted computer-use baseline: on AgentNetBench, it predicts 59.8% of recorded human actions exactly, compared with 53.7% for the prompted baseline, while on A11y-CUA it produces similar sequence fidelity (0.437 versus 0.465). In a single illustrative run example it also reproduces much of the relational structure among human-reported attributes such as workload, effort, frustration, and attentiveness ($ρs = 0.860$), although their magnitudes remain imperfectly calibrated. These results show that an explicit cognitive architecture can expose examinable and modifiable intermediate mechanisms while maintaining if not improving fidelity to human behavior