Abstract: As a crucial research direction in low-level computer vision, image restoration aims to reconstruct high-quality images from multiple degradations, encompassing denoising, dehazing, and super-resolution reconstruction. Traditional image restoration methods have evolved from specialized models to general models, which have improved generalization but still face significant limitations in aspects such as restoration fidelity and adaptability to cross-domain scenarios. In recent years, agent technologies—especially agent systems driven by large language models—have brought a novel solution to image restoration, leveraging their strong capabilities in cross-modal understanding, general reasoning, and natural language interaction. This paper systematically reviews the development of image restoration tasks and agent technologies, summarizes the evolution path from “specialized models” to “general models” to “agentic systems” in this field. It focuses on analyzing the cognitive architecture and core technologies of agents based on large language models, and proposes an intelligence level standard for image restoration agent systems. Finally, the paper discusses the challenges faced by agents in efficiency, generalization, quality assessment, cognitive architecture, and ethical security, and prospects future research directions including efficiency optimization and autonomous evolution, providing theoretical and practical references for the intelligent development of image restoration.