Skip to main content
Voice-agent APIs combine configuration with real-world calling capability. Keep agents in draft or inactive state until their prompt, voice, knowledge, tools, channels, and safeguards are ready.
Test-call and publish operations can create external side effects and should never be triggered by an unattended retry loop.

Configuration checklist

  • Persona and greeting are clear and safe.
  • System instructions define the objective and boundaries.
  • Knowledge sources are current and processed successfully.
  • Tools have the minimum required permissions.
  • Channel assignments point to the intended accounts.
  • Inbound configuration is explicit before enabling inbound traffic.
  • A human reviews a test call before publishing.
Use browser test calls for fast iteration and phone test calls only with an approved test number. Both are POST /calls with test: true. On timeout, inspect call history (GET /calls) before retrying. An agent’s settings, inbound behavior, channels, knowledge bases, tools and publish state are all changed with one call, PATCH /agents/{agent_id}. The steps run in a fixed order and, if one fails, earlier steps stay applied; the error lists them under applied. Read the agent (GET /agents?id=...) before you repeat the call.