Agents & publishing
An agent combines business instructions with the speech services that carry the conversation.
Choose an engine
| Engine | Configure in EVA | Best starting point |
|---|---|---|
| EVA Speech Engine | Instructions, recognition, language model, voice, and supported tools | A configurable agent you want to use through EVA’s API or dashboard |
| Cartesia hosted agent | A connection to a hosted agent you already own | An agent maintained in your Cartesia workspace |
A Cartesia agent must belong to the workspace accessible to the connected Cartesia credentials. An agent ID from another workspace will not work.
Configure an EVA agent
Set its instructions, greeting, language, recognition provider, language model, and voice. New EVA drafts use Deepgram recognition and a MiniMax language model by default. Existing agents keep their saved selections.
Use the selectable model and voice menus. Available choices come from a stored catalog; you do not need to query a provider on every page visit. Provider capabilities vary by language and model.
For natural conversations, keep the greeting short and ask one question at a time. Specify what the agent should do when it does not know an answer, and distinguish an offered action from a completed action.
Drafts and versions
- Edit and save the draft.
- Publish it when the configuration is complete.
- Test the published version.
- Select that version in your calling connection or API integration.
A published version cannot be edited. Publish a new version for changes, then update connections that should use it. Existing version IDs continue to identify their original configuration.
Tools and deletion
Only attach tools your business needs. Describe when a tool should run and how to handle a failure. Review outcomes before exposing an agent to callers.
Use the agent’s delete action to remove an agent you no longer need. Review any calling connections that use it first.
Short interruptions
The engine supports interruption handling, but transcription quality depends on the selected recognition provider and language. Test one-word phrases in your callers’ language. If a provider returns no transcript, the model cannot interpret that missing word; try a different recognition provider and compare the saved transcripts.