Reliability settings

Settings, Reliability controls what the voice runtime switches to when a provider fails during a call, plus two workspace limits. The page header reads: What the voice runtime switches to when a provider fails during a call, and workspace limits.

The page has, from top to bottom:

1. An Account state badge, with the reason when the state is not active.
2. Provider fallbacks: Fallback voice, LLM fallback model, LLM fallback after (ms) and Speech recognition fallback.
3. Limits: LLM token limit, Conductor pay-as-you-go and, when that is on, Conductor monthly limit (AUD).
4. A Save button that writes the whole page.

Settings apply to calls that start after you save. Calls in progress keep the settings they started with.

Provider fallbacks#

The dashboard offers gpt-5.6-terra, gpt-5.6-terra-mini, claude-sonnet-4-6 and gemini-2.5-flash as LLM fallback models. Through the API, llm_fallback_model accepts any model name.

How each fallback works#

Voice. The runtime builds a chain: the agent's voice, then the agent's own Fallback Voices (fallback_voice_ids, set in the agent's Security & Fallback Settings), then the workspace Fallback voice. When a voice errors, the next one in the chain speaks. The workspace voice is skipped if it is already the agent's voice or in its list.

LLM. The agent's model and the fallback model are wrapped together. If the agent's model errors, or sends no first token within LLM fallback after (ms), the same turn is retried on the fallback model. No fallback is built when the fallback model is the agent's own model.

Speech recognition. If the primary STT provider fails, the fallback provider takes over the stream. A non-Deepgram fallback needs that provider's API key on the runtime Needs keys; without it, no STT fallback is used and the runtime logs a warning. Choosing the agent's own provider has no effect.

When there is no fallback#

A non-recoverable pipeline error ends the call with error_tts, error_llm or error_asr (error_unknown for other errors). See Debug disconnections.

Choosing fallbacks#

• Pick a fallback voice from a different provider than your main voices, so one provider's outage does not take out both.
• Pick an LLM fallback from a different provider than the agent's model for the same reason, and check that it handles your prompts and tools in tests.
• Set LLM fallback after (ms) above your usual LLM p99 from Actual latency, or the fallback will fire on normal slow turns.

Limits#

The LLM token limit is described in Limits and concurrency.

Who can change what#

Provider fallbacks can be saved by admins and developers. The limits are admin only. Save writes both, so a developer should change only the fallbacks.

API#

curl -X PUT https://api.tried.com.au/workspace/reliability \
  -H "Authorization: Bearer $TRIED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "fallback_voice_id": "cartesia-Sarah",
    "llm_fallback_model": "claude-sonnet-4-6",
    "llm_fallback_timeout_ms": 3000,
    "stt_fallback_provider": "soniox"
  }'

{
  "fallback_voice_id": "cartesia-Sarah",
  "llm_fallback_model": "claude-sonnet-4-6",
  "stt_fallback_provider": "soniox",
  "llm_fallback_timeout_ms": 3000
}

Account state#

The badge at the top shows Active, Verification Needed, Invoice Past Due or Invoice Overdue. See Billing for what each blocks.