Skip to main content
A multilingual agent needs a language policy, reviewed translations, compatible voice settings, and tests in each language. Translating a few prompts doesn’t make the full experience multilingual. Decide which languages and channels you support, who owns each translation, and what happens when a language is unavailable. Keep the policy and operational facts in one authoritative source. Mark unreviewed or missing text clearly: the platform doesn’t translate missing locale entries automatically.

Keep the language choice explicit

For a text interaction, a channel can supply language and locale in its interaction context. In Studio’s test panel, set Language, Locale, and Timezone so the test matches the intended user. A flow can set a language after a person chooses one, or clear that choice to return to the default. Don’t depend on text-language detection, which isn’t supported.
ABL: language configuration
Locale files follow the configured fallback chain; for example, a regional locale can fall back to its base language. A missing key isn’t translated silently. Use the fallback text on purpose, and review what the person will hear or read. Templates are channel-specific, not language-specific. Keep the language choice in the interaction context and the localized strings, instead of creating a template variant for every language. See General settings for localization assets.

Test the whole conversation in each language

  • Translate and review every fixed prompt, error response, button label, and confirmation.
  • Write personas and opening messages in each supported language. Text eval personas and scenarios have no separate language field, so make the language clear in their content.
  • Set the Studio test interaction context to the language and locale you are checking.
  • Add a scenario where the person changes language mid-conversation. Check that the next response and any later tool-facing prompt use the selected language.
  • For voice, verify the configured ASR and TTS profile. Then test natural speech, local names, digits, negation, and mid-sentence correction.
  • Have a fluent reviewer read or listen to samples. A passing exact check can’t judge translation quality by itself.
For example, to add Spanish to Voyager: translate the cancellation read-back, choose Spanish in the test interaction context, and add a scenario that switches from English to Spanish. Language support spans the agent, locale assets, knowledge sources, channel, and eval set. Track it as one release capability, with an owner for each layer.