Wireless Agentic AI with Retrieval-Augmented Multimodal Semantic Perception

The rapid growth of multimodal agentic AI and LLMs enables richer perception and decision making, yet bandwidth-limited multi-agent links hinder timely exchange of task-critical semantics. Existing approaches based on static compression or heuristic selection do not adapt to the receiver's current query or to contention on a shared channel, which leads to redundant transmission or missing task-critical details. We propose Retrieval-Augmented Multimodal Semantic Communication (RAMSemCom), a receiver-driven framework that iteratively retrieves only high-resolution image patches most relevant to the active task when the initial downsampled summary is insufficient. A centralized Deep Reinforcement Learning (DRL) scheduler at the roadside unit coordinates per-agent patch budgets and timing under shared bandwidth and round deadlines, using semantic relevance and channel state to orchestrate concurrent requests. Retrieved patches are overlaid via a spatial-awareness layout to preserve scene geometry without full-frame transfer. In a multi-vehicle urban driving case study, RAMSemCom achieves earlier task completion and better performance than strong non-learning and token-level baselines under identical bandwidth and latency budgets.