Resource management is a fundamental challenge for distributed edge networks, and conventional reinforcement learning (RL) works well for single‑service scheduling but struggles with reward generalization in complex workflows. To address this, we introduce RAGAM‑AI, a hybrid framework that fuses RL algorithms with generative AI (Gen‑AI) to enable adaptive optimization of resource allocation at Multi‑access Edge Computing (MEC) units. The key idea is to employ a large language model (LLM) as a feedback provider, thereby mitigating reward generalization issues. Implementation-wise, a Multi‑Agent Soft‑Actor Critic (MASAC) is deployed with $$\Sigma$$tral serving as the reward oracle, handling fine‑grained scheduling of CPU, memory and bandwidth. The system first interprets global state information, then evaluates multi‑objective trade‑offs across distributed edge server agents to drive the resource orchestration loop. To reduce computational overhead, a surrogate reward model approximates the true reward function during training. Experiments compare RAGAM‑AI against single‑agent SAC, REINFORCE and random policies, showing that it achieves markedly lower allocation deviations for CPU, memory and bandwidth while keeping latency and data‑rate costs low. Overall, RAGAM‑AI offers a scalable AI‑as‑Feedback (AIAF) solution for efficient network resource optimization.\n\nReview: By closing the loop with LLM feedback and leveraging MASAC’s multi‑agent coordination, this work uniquely tackles the reward generalization problem in edge resource management, presenting a practical pathway for real‑world deployment.