The Agents screen in the TeleQuick console is the editor for one row of agent_config per agent. The left-hand list is every agent_config row that belongs to the active org, read directly from the database filtered on org_id and ordered by created_at. The first row is selected automatically when the screen loads. Selecting an agent opens the editor. The editor has three view modes and, inside the default one, four tabs: A page-level sub-nav switches between Agents (the pool plus this editor) and Voices (the org’s cloned-voice manager). Voices live here because they are an agent asset.

​
The agent pool and the four editor tabs

The config view is split into four panes — Conversation, Models & Voice, Actions, and Advanced. The split is presentation only. All four panes edit the same pending-edit object over the selected row, and one Save writes them together through adminAgents.updateConfig. Switching tabs never discards an unsaved change made on another tab. Per-agent actions in the pool: Every write goes through these org-scoped, audited BFF procedures rather than a browser database client, so an agent write is always attributed and always scoped to the org you are acting in. Creating an agent is gated by the org’s entitlement limit; the New Agent control is wrapped in the limit gate and will tell you when the pool is full.

​
Create an agent: name rules and templates

The agent name is a dispatch identifier, not a display label, so the modal enforces the charset its helper copy promises:
  • letters, digits and dashes only,
  • must start with a letter or a digit,
  • at most 64 characters,
  • required.
A bad name is reported when you submit, not while you type. The template picker pre-fills system_prompt and welcome_message only. It sets nothing else: New agents are created with coherent cascaded defaults: pipeline_mode is cascaded and turn_detection_type is silero.

​
What each tab covers

Per-agent credentials are columns on the same row — openai_api_key, anthropic_api_key, gemini_api_key, elevenlabs_api_key, deepgram_api_key and ollama_endpoint. They override whatever the org would otherwise use for that provider. ollama_endpoint must be an http or https URL. background_audio_url accepts s3, file, https or http.

​
Cascaded and realtime pipelines

pipeline_mode decides which half of the Models & Voice tab is meaningful.
  • cascaded runs three separate stages, so it reads the ASR, LLM and TTS provider/model selections and the TTS voice.
  • realtime runs a single speech-to-speech model, so it reads the realtime model and realtime voice lists instead.
The two modes pair with different turn detectors. cascaded is the right shape for any non-realtime model and defaults to silero; realtime sessions use the provider’s own server_vad. Changing pipeline_mode in the editor automatically flips turn_detection_type so the pair stays coherent — if you want a different detector, set it after changing the mode.

​
Turn detection fields and their accepted ranges

These ranges are enforced in the browser on Save because HTML min/max on a number input is bypassed by typing into it. Leaving a numeric field empty is valid: the server’s default applies. Semantic end-of-turn lives in the JSONB turn_detection_config column, not in its own columns. The editor exposes three of its params keys: eou_enabled, eou_threshold and eou_hold_ms.

​
Preserving turn_detection_config

turn_detection_config is a strict JSONB block with three top-level