The proliferation of diverse services in fifth generation (5G) and beyond networks necessitates dynamic, fine-grained resource management to guarantee strict Quality of Service (QoS) for mission-critical applications. Traditional static resource allocation models fail to address the highly variable traffic demands and heterogeneous requirements inherent in network slicing (NS) within a Software-Defined Networking (SDN) framework. This paper proposes a novel framework utilizing Deep Reinforcement Learning (DRL), specifically a Multi-Agent Deep Deterministic Policy Gradient (MA-DDPG) approach, to achieve autonomous and optimal resource orchestration for dynamic NS in critical telecom infrastructure. Our proposed DRL agent learns a real-time mapping between the evolving network state (e.g., slice demand, available resources, and congestion) and optimal resource allocation decisions (e.g., CPU, memory, and bandwidth allocation across Virtual Network Functions, VNFs). Simulation results, comparing the MA-DDPG approach against established benchmarks like Greedy and traditional Deep Q-Networks (DQN), demonstrate a significant improvement in Service Level Agreement (SLA) violation rate reduction, resource utilization efficiency, and slice admission control performance. This DRL-driven approach is critical for the reliable and efficient operation of future resilient, ultra-low-latency critical communication systems