This work develops a hybrid framework for reservoir planning in Mendoza, integrating multicriteria GIS analysis, multi-agent simulation originally prototyped in SIG-MAS (Python) and fully implemented in Python 3.10+, and extensive temporal validation against SARIMA and Naive Seasonal statistical models. The approach combines spatial selection of optimal sites at 30‑meter resolution, hydrological forecasting based on 67 years of data, and behavioral simulation with heterogeneous agents, resulting in a calibrated, spatially explicit simulation model that operates in batch mode using historical data, without IoT integration or real‑time updating. The results show that, for the 2019–2023 period, the SIG-MAS (Python) model delivered the best performance in the Mendoza River basin—characterized by high variability (CV ≈ 0.65)—with an RMSE of 14.6 (valor puntual estimado) m³/month and an NSE of 0.830, outperforming both the calibrated SARIMA model (RMSE = 18.90; NSE = 0.755) and the Naive Seasonal model. In the Tunuyán River, a low‑variability basin (CV ≈ 0.35), the Naive model achieved the best performance, with an RMSE of 8.78 m³/month and an NSE of 0.823, followed by calibrated SARIMA and then SIG-MAS (Python). Local SARIMA calibration eliminated the initially observed negative NSE values and reduced errors by up to 60% compared with its generic version, highlighting the importance of basin‑specific tuning. The implemented architecture corresponds to a digital simulation fed with historical data in batch mode, without IoT integration or real‑time updating. Recalibration of parameters and the removal of the guardrail are recommended for future versions. The proposed approach is replicable and contributes to decision‑making in semiarid regions, although improvements are needed for extreme‑event detection and operational validation.