Skip to main content
Voice agents combine configuration with real-world communication. Treat every test-call, publish, inbound-channel, and phone operation as a controlled side effect.

Build sequence

  • Create the agent with POST /agents and draft: true.
  • Update the agent prompt, voice, greeting, and safeguards with PATCH /agents/{agent_id}.
  • Attach processed knowledge bases and only the tools the agent needs, with knowledge_base_ids and tool_ids on the same update.
  • Configure channels and inbound behavior explicitly, with channels and inbound on the same update.
  • Use browser test calls with synthetic data first: POST /calls with test: true, mode: "browser" and the agent_id.
  • For browser microphone setup and the website widget path, see WebRTC browser calls.
  • Use phone test calls only with an approved test number: POST /calls with test: true, mode: "phone", the agent_id and test_phone.
  • Publish after human review of the transcript and tool behavior.

Test calls

Test calls may consume credits, create call records, and contact a real number. They are not safe to repeat automatically after a timeout. Query call history or the returned job/call ID before retrying.

Publishing

Publish with PATCH /agents/{agent_id} and publish: true. Publishing needs a confirmation_id: a short-lived confirmation issued for the agent’s current configuration by the Sendora assistant connection when it validates that configuration. It is tied to your workspace and to a configuration that has not changed since, so editing the agent invalidates it (409, confirmation_invalid). Publishing changes which configuration can receive traffic. Keep a copy of the reviewed configuration, record the agent public ID and publish result, and disable or update channels deliberately if the agent behaves unexpectedly.