Large Language Model-Based Translation Agents: A Review and Future Perspectives

Large language models (LLMs) have enabled general-purpose agents to expand into vertical domains such as translation. Accordingly, translation tools are evolving from passive programs into translation agents capable of autonomous planning, memory management, and external resource use. This paper reviews recent studies on translation agents from the perspective of Translation Studies rather than from a purely algorithmic viewpoint. It traces their technological evolution and theoretical mapping, with particular attention to single prompting with retrieval-augmented generation, multi-agent collaboration, and multi-level memory mechanisms. These developments partly simulate and reconstruct the cognitive strategies and collaborative networks of human translators. At the same time, the integration of translation agents into complex workflows reveals persistent problems, including limitations in quality assessment, blurred ethical accountability, and fluctuations in translators’ cognitive load. The paper argues that future research should further examine evaluation mechanisms, ethical norms, and responsibility allocation within human-AI collaborative translation while also strengthening translator prompt literacy, digital resilience, and curriculum reform in translation education. The significance of translation agents lies not in replacing human translators, but in reshaping the translation ecosystem and its collaborative models.