Teaching AI to translate like a professional communicator

Current machine translation systems treat translation as text-in/text-out conversion, ignoring the communicative intent behind it. This prototype introduces an agentic cycle where users first compose a structured translation brief through dialogue—specifying skopos, register, audience, and genre—then the system generates and verifies output using evidence-grounded error scoring. The system maintains document coherence via named-entity memory and bilingual summaries. This is a conceptual and architectural contribution: the authors argue translation in the generative AI era should be communication design, not mechanical conversion. No empirical validation is included; the work is grounded in translation studies theory.