Skip to main content
Choose a language model for the work your agent needs to do: understand the request, follow the prompt, and produce an appropriate response. Start with the current model catalog so the identifier and available settings match Threetone.
Creation and draft saving are separate from publication. Later publication checks model availability and compatibility. Managed text-only configurations can pass those checks; voice and custom-model publication are currently unavailable. See current availability.

​
Choose and review a model

1

Check the catalog

Read Models and List available LLMs. Choose an available managed model and keep its exact identifier. A provider’s model name is not necessarily its Threetone catalog identifier.
2

Set the primary model

In the agent editor, open LLM to review the model and its settings. The saved primary model is conversation_config.agent.prompt.llm.For API authoring, read the intended branch first and preserve the rest of its configuration. Use an explicit branch_id when updating conversation settings.
3

Review supported settings

Set only options supported by the chosen model. Use the table below to locate them, and review LLM cascading before selecting backups.
4

Save and read back

Save a complete server draft. If you publish an eligible managed text-only configuration, review the diff and record the new version ID. Read the saved document with Get agent and check the primary model and backup policy together.

​
Generation settings

These fields belong to conversation_config.agent.prompt. A successful save is evidence of configuration storage. It does not measure response quality, latency, or successful failover. Prepare a small set of representative prompts before comparing models when execution is available.

​
Frequently asked questions

Creation stores the initial authoring configuration. Later publication also checks available models and compiles an execution configuration. Check the primary model, backups, and whether the agent is text-only.
You can prepare a custom model configuration, but custom-model publication is unavailable. An endpoint probe does not remove that restriction.