Joint SLA-Aware Task Offloading and Adaptive Service Orchestration With Graph-Attentive Multi-Agent Reinforcement Learning

Coordinated service offloading is essential to meet Quality-of-Service (QoS) targets under non-stationary edge traffic. Yet conventional schedulers lack dynamic prioritization, causing deadline violations for delay-sensitive, lower-priority flows. We present PRONTO, a multi-agent framework with centralized training and decentralized execution (CTDE) that jointly optimizes SLA-aware offloading and adaptive service orchestration. PRONTO builds on Twin Delayed Deep Deterministic Policy Gradient (TD3) and incorporates spatiotemporal, topology-aware graph attention with top-K masking and temperature scaling to encode neighborhood influence at linear coordination cost. Gated Recurrent Units (GRUs) filter temporal features, while a hybrid reward couples task urgency, SLA satisfaction, and utilization costs. A priority-aware slicing policy divides bandwidth and compute between latency-critical and throughput-oriented flows. To improve robustness, we employ stability regularizers (temporal smoothing and confidence-weighted neighbor alignment), mitigating action jitter under bursts. Extensive evaluations show superior QoS and channel utilization, with up to 27.4% lower service delay and over 18% higher SLA Satisfaction Rate (SSR) compared with strong baselines.