Skip to main content

Voice Agents

Voice Agents are a dedicated voice-first surface in SimplAI. They have their own create flow, list view, configuration screen, version history, run history, and tracing — separate from text-based Agents because voice has fundamentally different concerns: continuous bidirectional audio, an STT → LLM → TTS pipeline (or the OpenAI Realtime API), and strict sub-second latency budgets.

When to build a Voice Agent (vs. a text Agent)​

Use a Voice Agent when:

  • The end user speaks and listens — phone, mobile mic, web voice widget.
  • You need to manage interruption / barge-in.
  • Latency to first speech matters more than richness of output.
  • You want telephony integration (inbound calls and outbound dialing).

Use a text Agent when the surface is chat, email, or any text-based UI. Some text-agent features — Browser Use, Artifacts, Reflection, Follow-up Messages — are not available to voice agents because they don't translate to a voice interaction.

What this section covers​

  • Creating a Voice Agent — the dedicated nav, list view, and create flow.
  • Configuration — every setting on the voice agent configuration screen: execution modes, LLM, STT, TTS, voice cloning, multilingual, VAD, BYOK, tools, knowledge bases.
  • Deployment — JS SDK, embed widget, and voice calling tabs.
  • Voice Calling (telephony) — inbound number setup and outbound dialing.
  • Run history & traces — call recordings, conversation transcripts, span-level latency, and audio playback synced to the trace.
  • Voice Agent as a workflow node — invoking a voice agent from a Tool or another agent.
  • Evaluation — running evaluators against voice transcripts.
  • Pricing & billing — per-second metering, platform-managed vs. BYOK breakdown.
  • Latency — TTFS, P50/P90/P99 metrics, and the latency dashboard.

Things to keep in mind​

  • Voice Agents and standard Agents are different resources. A voice agent never appears in the Agents list, and vice versa.
  • Voice runs use either the Sync or Async path, or the OpenAI Realtime API path. Sync mode bypasses queues entirely; Async runs through a dedicated voice priority queue isolated from batch workloads.
  • Tools attached to voice agents must be in Sync mode so they can return inside the latency budget. The platform enforces this when you attach tools.
  • Voice Agents are billed per second of call duration, plus tool/KB charges and connection charges. See Pricing.