To assess the progress in terminology-aware machine translation, the Terminology Shared Task is hosted alongside the Conference on Machine Translation (WMT). In this paper, we describe our submission to the shared task: STaR-MT. This is a three-step pipeline composed of (S)election, (T)ranslation, (a)nd (R)evision, focused on terminology-aware machine translation (MT). Our methodology adopts a lightweight agent-based design that selects relevant entries or examples and presents them to the translation agent, which uses a general-purpose LLM with a large context window to produce the initial output. Next, the revision agent uses language-specific models to improve translation quality for selected target languages. Experiments on the FLORES+ dataset with different model sizes in the pipeline suggest that machine translation pipelines can benefit from combining large general-purpose translation models with smaller, language-specific revision models, at least in some language pairs. While simple in architecture, this approach allows us to generate translations at the document level, with no need for pre-processing or paragraph splitting, and with minimal errors across most languages.