Agentic AI Translate: An Agentic Translator Prototype for Translation as Communication Design

The paper's core contribution is treating translation theory (skopos, register, audience) as formal constraints that shape agentic behavior, not as post-hoc evaluation criteria. This inverts how specialized knowledge typically enters AI systems: instead of training on examples and hoping the model absorbs domain logic, the researchers encode theory directly into the agent's decision loop.

This connects directly to RAGA's approach from the same day. Both papers embed verification and structured reasoning into agentic workflows rather than relying on end-to-end learned behavior. Where RAGA uses a Read-Search-Verify-Construct loop to audit knowledge graph assembly, Agentic AI Translate uses a four-stage workflow grounded in translation briefs and error protocols. The shared pattern is treating domain expertise as executable specifications that make agent decisions auditable and reproducible, not opaque. This is distinct from the coding agent benchmark (1GC-7RC) which measures whether agents can do specialized work, whereas these two papers focus on how to formally encode what "correct" work looks like.

If Yamada or Kocmi publish follow-up work applying this theory-as-specification pattern to other translation pairs or language families within the next 12 months, that signals the approach generalizes beyond the prototype. If the GEMBA-MQM validation protocol becomes adopted in commercial MT evaluation, that confirms the field is moving toward mechanistic error frameworks rather than holistic quality scores.