<![CDATA[Load balancing in bare-metal server environments presents a unique challenge, as conventional static algorithms and single-agent Deep Reinforcement Learning (DRL) often fail to adapt securely to extreme traffic variations. Single DRL agents face a convergence dilemma: Advantage Actor-Critic (A2C) suffers from policy degradation under sudden stress, while Proximal Policy Optimization (PPO) lacks rapid responsiveness in low-traffic states. To address these limitations, this study proposes a Hybrid Adaptive Load Balancing (HALB) architecture that dynamically integrates A2C and PPO policies using hysteresis switching logic, fortified by a deterministic Safety Shield. The system was evaluated using an Nginx reverse proxy against standalone A2C and PPO baselines under Ramp-up, Sudden Spike, and High Traffic scenarios. Experimental results demonstrate that HALB provides noticeable improvements over single-policy agents, particularly in latency optimization and tail latency truncation. During extreme sudden traffic spikes, the HALB system constrained the 95th Percentile (P95) latency to 77.33 ms, compared to standalone A2C (85.00 ms) and PPO (85.23 ms). Under sustained high traffic conditions, HALB achieved an overall average latency of 23.64 ms while maintaining a load distribution fairness index (J > 0.99). Fortified by the Safety Shield that prevented node crashes at an 85% CPU threshold, the HALB architecture offers a balanced approach between intelligent responsiveness and operational stability. This research provides a resilient traffic management framework for rigid infrastructures, contributing to the development of safe adaptive control in non-elastic distributed systems.]]>
