<p>Heating, ventilation, and air conditioning (HVAC) systems consume a significant amount of energy to maintain thermal comfort and indoor air quality in buildings, which results in high operational costs. Reinforcement learning is an effective method for controlling HVAC systems. However, in large and complex HVAC systems, traditional reinforcement learning algorithms often face the challenges of slow training speed and poor convergence performance. This paper proposes a multi-objective optimization control method based on the multi-agent deep deterministic policy gradient (MADDPG) algorithm, which aims to minimize HVAC energy consumption while ensuring optimal thermal comfort and indoor air quality in each zone. Using a multi-zone office building with fan coil units and a dedicated outdoor air system as a case study, we developed an EnergyPlus-Python co-simulation platform. The proposed control method was employed during both the heating and cooling seasons to independently control the temperature setpoints and fresh airflow in different zones of the office building. The simulation results from both the heating and cooling seasons demonstrate that the MADDPG control method exhibits faster convergence during training and excellent learning capabilities, allowing it to adapt effectively to changes in environmental conditions and implement appropriate control actions. Under similar indoor thermal comfort and air quality conditions, the MADDPG control method consumes less energy than the traditional reinforcement learning method, it saves 24.1% of energy during the heating season and 8.9% during the cooling season compared to the rule-based control method. Additionally, by adjusting the reward function in the MADDPG algorithm, it is possible to flexibly balance energy consumption, thermal comfort, and air quality preferences, demonstrating the algorithm’s strong applicability.</p>