Agentic Tool-Augmented Translation for Low-Resource Sanskrit--English under Resource Constraints

Machine translation for low-resource and morphologically rich languages remains challenging due to limited parallel data and high computational barriers, particularly for classical languages such as Sanskrit. We propose a training-free, agentic, tool-augmented decoding framework that grounds generation in external linguistic resources, including dictionary lookup, morphological and syntactic analysis, dynamic example retrieval, and glossary constraints. The agent orchestrates these tools at inference time to improve lexical disambiguation, grammatical accuracy, and terminology consistency without any parameter updates. Experiments on five Sanskrit--English benchmarks across diverse domains show consistent gains over a tool-free baseline, with the full system achieving the best BLEU and chrF scores among all variants and approaching stronger external baselines despite using a much smaller model. These results demonstrate that tool-augmented, agentic methods provide an effective and accessible solution for low-resource translation in resource-constrained settings.