Recording pending: Configure an agent. Follow the written guide below.
The dashboard submits prompt, first-message, LLM, soft-timeout, and dynamic-variable edits through the configuration publication flow. Publication currently requires an eligible managed text-only configuration; voice and custom-model publication are blocked. Its Voices, Language, ASR, and Agent behavior pickers are previews whose changes are not saved; the Security page also displays disabled example settings. Use a complete server draft to preserve those settings without publication. See voice and language and conversation flow.
Configuration structure
Configure and review
1
Set the purpose and opening
Define the task in
agent.prompt.prompt and write a matching agent.first_message. Keep the first message short enough to establish the purpose without asking several questions at once.2
Choose language and models
Set
agent.language and review agent.response_language_policy. Select an available LLM with agent.prompt.llm.For voice configurations, include ASR and TTS settings with compatible models and a saved voice. Check the catalog rather than assuming a provider supports every language and recognition mode.3
Attach workspace resources
Review each tool, MCP server, and knowledge attachment. Use the intended identifiers and check access in the active workspace.Keep secrets in supported credential storage. Resource identifiers and prompt instructions do not replace authorization.
4
Review the entire version
Inspect the configuration diff, including
platform_settings and workflow. Preserve unrelated fields when preparing a full draft snapshot.Save a complete server draft on the intended branch and read it back. Only publish when the configuration is eligible; managed text-only publication has model, resource, and workflow checks. See versions and drafts.conversation_config. The generated Create agent, Get agent, and Update agent pages define the request and response shapes.
Configuration and execution
The documented implementation currently gates signed-token session initiation, outbound calling, traffic deployment, and automated response-test invocation. A valid configuration can be saved and reviewed without proving that those execution paths work. See current availability.
Troubleshooting
Start with Create your first agent for the dashboard workflow or API quickstart for a server-side request example.
Frequently asked questions
Do resource IDs grant access to the resource?
Do resource IDs grant access to the resource?
No. Each referenced voice, tool, MCP server, or knowledge source must be available to the intended workspace and caller. Inspect access separately from the configuration reference.
How do I confirm which settings were saved?
How do I confirm which settings were saved?
Read the intended branch or version, compare its complete configuration with the reviewed diff, and retain the version ID. A saved setting does not prove that its runtime behavior is available or enforced.