Every agent is a single config document. It names the pipeline shape, the provider and model for each stage, how the runtime decides when the caller has finished talking, and the budgets that bound a call. You edit it in the console at agent.telequick.dev or push it through the control-plane API; the gateway hydrates it into the runtime at call setup and hot-reloads tunables without dropping in-flight calls. There is no rebuild and no redeploy — save the document, place a call. This page is the field-by-field reference. If you’re wiring a pipeline for the first time, start with BYO ASR/LLM/TTS (cascade) or BYO speech-to-speech (duplex) and come back here to tune the knobs.

​
The shape of agent-config.yml

A cascaded agent — the most common shape — looks like this. Every field below has a default, so a minimal config is just a name, a system prompt, and any providers you want to override.
agent-config.yml
name: support-agent
pipeline_mode: cascaded           # cascaded (ASR → LLM → TTS) | realtime (duplex)

# --- Models & providers ---
asr_provider: deepgram
asr_model: nova-2
llm_provider: openai
llm_model: gpt-4o-mini
temperature: 0.4
conversation_memory_messages: 20
llm_max_retries: 2
llm_fallback_provider: anthropic
tts_provider: elevenlabs
tts_voice: <voice-id>
language: en
system_prompt: >
  You are a friendly support agent for Acme. Keep replies short and
  spoken-friendly.

# --- Turn detection ---
turn_detection:
  type: local                     # local (on-device VAD) | server_vad
  silence_threshold_ms: 500
  min_speech_ms: 300
  prefix_padding_ms: 200
  barge_confirm_ms: 300           # hold-and-confirm barge-in window
  response_min_speech_ms: 600

# --- Budgets & guardrails ---
max_session: 15m                  # hard ceiling — the runtime hangs up at this
max_output_tokens: 4096
idle:
  soft_prompt_s: 20               # re-engage prompt after this much silence
  hangup_s: 60                    # then end the call
  soft_prompt_text: "Are you still there?"
A duplex agent replaces the cascade fields with a single realtime node — see Realtime (duplex) agents below. Everything in the turn-detection and budgets sections applies to both shapes.

​
Models & providers

Pick the pipeline shape with pipeline_mode, then fill in the stage fields for that shape. The runtime processes audio as PCM16 LE at 8 kHz end to end, so every provider is fed and drained at that rate — providers that natively want 16/24 kHz are resampled for you.
​
pipeline_mode
string
default:"cascaded"
cascaded chains three provider sessions (speech-to-text → language model → text-to-speech). realtime hands the whole turn to one duplex speech-to-speech model. Choose the cascade for per-stage control and the widest model choice; choose realtime for the lowest turn latency.
​
name
string
required
A human-readable agent name. Shown in the console and stamped onto call records so you can tell agents apart in the dashboards.
​
system_prompt
string
The agent’s persona and task. Keep replies “spoken-friendly” — short sentences read better than paragraphs over a phone line.
​
language
string
default:"en"
ISO 639-1 hint. Prepends a hard directive to the system prompt so the model answers in your market’s language instead of mirroring the caller.

​
Cascade fields (pipeline_mode: cascaded)

​
asr_provider
string
default:"deepgram"
Streaming speech-to-text vendor. Supported ids include deepgram and elevenlabs. Full table on BYO ASR/LLM/TTS.
​
asr_model
string
default:"nova-2"
Forwarded to the ASR vendor as its model hint (Deepgram opens the socket with 8 kHz linear16).
​
llm_provider
string
default:"openai"
Language-model vendor: openai, anthropic, gemini, or ollama (self-hosted, no key). This is the stage where tool calling runs.
​
llm_model
string
Vendor model id, e.g. gpt-4o-mini, claude-sonnet-4-5, gemini-2.5-flash, llama3.1.
​
temperature
number
default:"0.4"
Sampling temperature for the language model.
​
conversation_memory_messages
integer
default:"20"
Rolling history the runtime replays into each turn. Larger keeps more context at the cost of prompt size and latency.
​
llm_max_retries
integer
default:"2"
Retries against the primary language-model provider on a failed turn before falling back.
​
llm_fallback_provider
string
A different provider entirely, tried once after llm_max_retries is exhausted. A resilience path (degrade one turn instead of dropping the call), not a load balancer.
​
tts_provider
string
default:"elevenlabs"
Text-to-speech vendor: elevenlabs, cartesia, deepgram, or openai. Streaming providers emit audio as LLM text arrives, which keeps first-audio low.
​
tts_voice
string
The provider’s voice id — an ElevenLabs voice_id, a Cartesia voice UUID, or a Deepgram Aura voice such as aura-asteria-en.

​
Realtime (duplex) agents

Set pipeline_mode: realtime and describe a single realtime node instead of the three cascade stages. There is no separate ASR or TTS to configure — one model listens and speaks on the same bidirectional session.
agent-config.yml
name: support-agent
pipeline_mode: realtime
entry_node: REALTIME
pipeline:
  REALTIME:
    node_type: REALTIME
    provider: openai
    model: gpt-realtime
    voice: shimmer            # alloy | shimmer | …
    instructions: |
      You are a friendly support agent for Acme. Keep replies short.
Cloud duplex providers (OpenAI Realtime, Gemini Live, xAI Grok Voice) are shipped. Serving your own open-weights duplex model on your GPUs is in preview — see self-hosted inference.

​
Credentials

Provider keys are per-tenant, never written into agent-config.yml. Set them in the console at agent.telequick.dev; they are resolved from the control plane at call setup, sealed at rest, and injected into the pipeline. An agent whose selected providers have no usable key is refused at session creation with a clean error rather than starting and failing mid-call. Environment-variable fallback for single-tenant and local development is covered under Credentials.

​
The turn_detection block

Turn detection is what makes a call feel natural — the agent stops the instant the caller barges in and replies the instant they finish. It has two parts: a detection mode (who runs voice activity detection) and a handful of timings. Start with the defaults and adjust from there. The concepts page, Turn Detection & Barge-In, explains the behavior end to end; this is the field reference.
​
turn_detection.type
string
default:"local"
local runs an on-device VAD on the gateway — the default, and the right choice for cascaded pipelines. server_vad defers end-of-turn detection to a cloud realtime provider that already emits speech-start/stop events; the runtime auto-selects it for a realtime entry node unless you set type explicitly. A cascade with server_vad never sees a turn boundary — use local there.
​
turn_detection.silence_threshold_ms
integer
default:"500"
Minimum trailing silence before end-of-utterance fires. Lower (300) for snappier replies; raise (800) for callers who pause to think. Below ~300 the agent starts interrupting on inter-word pauses.
​
turn_detection.min_speech_ms
integer
default:"300"
Discard candidate utterances shorter than this. Raise toward 500 if line noise or clicks cause false triggers.
​
turn_detection.prefix_padding_ms
integer
default:"200"
Audio kept before the speech-start marker. Raise if the first syllables are getting clipped on the way into the ASR.
​
turn_detection.silence_floor_dbfs
integer
default:"-45"
Signal level (dBFS) below which the frame is treated as silence. Quieter trunks want a lower floor (-50); noisy ones a higher one (-40).
​
turn_detection.barge_confirm_ms
integer
default:"300"
The hold-and-confirm barge-in window. A speech onset arms a pending barge; it only cancels the agent if speech sustains past this window — so a backchannel (“mhm”, “ok”, “yeah”) hits silence first and never cuts the agent off. Set to 0 for legacy immediate barge-in.
​
turn_detection.response_min_speech_ms
integer
default:"600"
Minimum speech duration before the caller’s turn is allowed to trigger a response. A coarser duration filter than min_speech_ms, applied to whole turns rather than candidate frames.
​
turn_detection.margin
number
Sensitivity margin on the detection threshold while the agent is idle.
​
turn_detection.margin_while_tts
number
Sensitivity margin while the agent is speaking. The telephony-safe default is equal to margin — the gate is not loosened during agent speech, because there is no echo cancellation in the media path and a loosened gate would let the agent’s own voice self-trigger a barge. Only lower it if you have echo-cancelled audio.
Backchannel suppression is acoustic only — utterance duration plus the hold-and-confirm window. Lexical gating (distinguishing “stop” from “ok”) is not built; plan around the acoustic behavior. See Turn Detection & Barge-In.

​
Codecs

Codec selection is negotiated per transport at call setup, not set in agent-config.yml. The runtime works in a single 8 kHz PCM16 contract internally, so the same agent behaves identically no matter what the caller’s leg negotiates.
Caller legWire codecInside the runtime
Phone (PSTN / SIP trunk)G.711 µ-law (PCMU) / A-law (PCMA) at 8 kHzdecoded to PCM16
Browser / mobile appOpus over QUICdecoded and resampled to PCM16
Opus runs narrowband for telephony — the encoder’s internal rate is 8 kHz to match the runtime contract, even though its RTP payload keeps the standard opus/48000/2 rtpmap and 48 kHz clock (a 960-tick advance per 20 ms frame). A 20 ms tick is 160 samples at 8 kHz. Audio is 8 kHz mono end to end; there is no video codec in the voice path. Details and how to influence carrier negotiation are on Codecs.

​
Session & output budgets

These bound how long a call can run and how much a single turn can cost. The runtime enforces them; you don’t have to build a watchdog. The Session Lifecycle page covers the full state machine.
​
max_session
duration
default:"15m"
Hard ceiling on total call duration. The runtime ends the session when it’s reached — a stuck or runaway agent cannot hold a line open indefinitely. This is a ceiling, not a target; most calls end well before it.
​
max_output_tokens
integer
default:"4096"
Cap on the language model’s output per response. Bounds token spend on any single turn (previously unbounded). Lower it for terse agents; keep headroom if the agent reads back long confirmations.

​
Idle watchdog

The idle watchdog re-engages a silent caller and then hangs up if they’ve truly gone. It is on by default. The same knobs apply to cascade and duplex agents.
​
idle.soft_prompt_s
integer
default:"20"
Silence, in seconds, before the agent emits a re-engage prompt (“Are you still there?”). Fires once per idle stretch.
​
idle.hangup_s
integer
default:"60"
Silence, in seconds, before the runtime ends the call. Measured from the start of the idle stretch, so it should exceed soft_prompt_s.
​
idle.soft_prompt_text
string
The text spoken (or synthesized) as the re-engage prompt. Leave unset for the runtime default.
Connect, idle, and per-stage turn timeouts also exist as operator-tunable runtime settings and are hot-reloaded without dropping calls — in-flight sessions keep the timeouts they captured at creation time, so a reload never changes the rules mid-call. If you run self-hosted, these are documented with the deployment models.

​
How a config change reaches a live call

1

Edit the config

Change agent-config.yml in the console at agent.telequick.dev or via the control-plane API.
2

Save — no rebuild

The config is hydrated into the runtime from the control plane. There is nothing to compile or redeploy.
3

New calls pick it up immediately

The next call routed to the agent uses the new config. Runtime tunables (timeouts and the like) hot-reload for the process; in-flight calls keep the config they started with so a mid-call edit can’t destabilize an active conversation.
4

Verify on a test call

Place a call and watch the latency breakdown and call traces to confirm the new providers, timings, and budgets took effect.

BYO ASR / LLM / TTS

Build a cascade — provider tables, models, and credentials.

BYO speech-to-speech

Run a single duplex model for the lowest turn latency.

Turn detection & barge-in

The behavior behind the turn_detection block.

Session lifecycle

From connect to teardown, and the guardrails that bound it.

Tool calling

Give the LLM stage HTTP, MCP, and client tools.

Codecs

What gets negotiated on the wire, and how to influence it.