<p>This paper presents a novel data-driven optimization framework for efficient integration of photovoltaic (PV) agents in residential microgrid systems. Using a multi-agent system architecture composed of software and physical agents implemented on Raspberry Pi boards, the proposed framework addresses the specific constraints of residential microgrid environments. Each agent autonomously controls components such as PV panels, storage systems, and household loads, with communication protocols and decision-making algorithms enabling seamless data exchange and coordination. A key innovation is the implementation of a PV agent using a data-driven modeling approach enhanced by a nonlinear autoregressive neural network with exogenous inputs (NARX) to predict PV energy production. The PV agent determines the optimal control mode, selecting between maximum power point tracking (MPPT) and limiting control, based on real-time conditions such as irradiation levels, outdoor temperature, home energy consumption, and battery state of charge (SOC). This strategy avoids excessive energy injection into the grid, ensures power quality, and maximizes energy capture. The effectiveness of the framework is validated both by simulation in a MATLAB Simscape environment and by real-world experiments with a laboratory-based platform. Performance metrics, including low mean absolute error and high correlation coefficients between predicted and measured PV power, demonstrate the predictive accuracy and reliability of the PV agent. The results highlight the potential of this approach to optimize energy distribution and improve the sustainability of residential microgrids.</p>