Skip to main content

Agent settings

Agent settings are available in the Settings tab of the agent builder. Open an agent, select Settings, and configure the options that control conversation handling, the agent’s speech, knowledge base usage, and call control. There are a lot of them, so they are split across separate pages — listed below. Here is what has to be decided first and a map of the tab.

Pipeline mode

The Pipeline mode card at the top of the tab has two options:
  • STT · LLM · TTS — configure separate speech-to-text, language model, and text-to-speech options.
  • Realtime model — use a real-time speech-to-speech model for the voice agent.
Pipeline mode toggle: STT-LLM-TTS or Realtime model
Decide this before the rest: some settings do not work in realtime mode. Such an agent has no separate synthesis step and does not fill fields as the conversation goes, so agent objectives, phrase corrections and pause markup have no effect. Objectives can still be switched on but do nothing — wherever that is the case, the page says so plainly. What only the STT · LLM · TTS pipeline uses is hidden: the Voice card is replaced by the Realtime card, the Execution model field disappears, and Additional no longer shows the classifier, transition, warm-up and speech sections.

What the Settings tab looks like

From top to bottom: Quality & integrations is a card with four rows. Each row shows a status — for example, On · 3 fields, the number of observer rules or No endpoints yet — and opens a window with that section. The window has a list of sections on the left, so you can switch between them without closing it.

Additional

The Additional row opens a window with all rare settings on one scrolling page. The sections are cards grouped by topic, with navigation on the left:
  • Access — Hide from members.
  • Models & routing — Request understanding, Transition classifier, LLM warm-up, Flex mode.
  • Fallback — Reserve services.
  • Speech — Pre-TTS normalizer, Turn detection (end of turn).
  • Operations — Farewell on platform shutdown, No media detection and Outbound SIP headers.
The Find a setting field above the navigation hides the cards that do not match. A section that is off stays collapsed until you turn it on with the switch in its header. In the navigation a dot marks a section: on — its switch is on, changed — a section without a switch holds a value that differs from its default. The Additional row on the tab shows a N changed badge: how many sections are on or differ from their defaults. Hide from members carries the Applies immediately tag: it is written at once, and Cancel does not undo it.

Two common controls

  • The gear next to a field or card opens that control’s parameters. You edit a copy: Save writes it to the draft, Cancel drops it. A dot on the gear means a value inside differs from its default. The Quality & integrations and Additional windows work the same way: edits from all their sections are saved together; if you close the window with unsaved edits, it asks Discard changes?.
  • «?» next to a field shows its help.
Everything you save goes to the agent’s draft. Publishing is a separate step.

Where each setting lives

Prompt, models and knowledge

The global prompt and variable substitution in authored text, execution and classifier models, the transition classifier, service tier, cache warm-up, knowledge bases, flex mode.

Voice: STT, TTS and VAD

Speech recognition and synthesis, turn-taking and interruptions, active listening, background audio, turn detection, the pre-TTS normalizer.

Answering machine detection

How the platform tells a person from a voicemail, which checks to turn on, what the agent does after the decision.

Phrase corrections and pauses

The organization’s replacement dictionary and [pause N] markup: how the agent sounds when it reads numbers and amounts.

Recording and call control

Call recording and its retention, the duration cap, silence reminders, the transfer failure message.

Agent objectives

What has to be established in a conversation, how the platform marks it done, and what the mechanism deliberately does not do.

Post-call analytics

Post-call analysis: summary, sentiment, success evaluation, custom fields, the deep tier.

Observers

Detectors of conversation events, the agent’s soft and hard reactions, the decision backend.

Webhooks

Signed notifications about the agent’s calls to your endpoints: events, secret, timeout.