Skip to main content

​
Pick a voice-agent model preset

A voice agent is three provider choices yelled at each other: which speech-to-text engine hears the caller, which LLM thinks, and which text-to-speech voice answers. The Model Presets picker (/agents/model-presets) bundles those three into one curated configuration per card, so you pick a latency-cost-quality tradeoff and get a working draft agent in one click instead of hand-tuning provider dropdowns. Every preset ships the same dependable foundation — Deepgram transcription and a Cartesia voice — and the presets differ only on the model tier and the lane it runs on. What you are choosing is the reasoning tier.

​
1. The four presets

Each picker card shows the comparison that matters up front: estimated time-to-first-audio latency (p50), estimated blended STT+LLM+TTS cost per minute, and a 1–5 reasoning-quality rating. Open Preview on a card for the full description and its recommended-for list. The numbers are curated comparison estimates — a selection aid, not per-call SLAs. Use them to choose a tier, then measure your own traffic on the VAQI dashboard once the agent is live. Balanced is the recommended starting point: Sonnet-tier reasoning at a mid-market per-minute price. If you do not have a specific constraint pushing you elsewhere, start here and re-tier up or down as VAQI tells you. High Intelligence trades latency and cost for the deepest reasoning Orbit runs. Pick it when a wrong answer costs more than 500 ms of hold time: collections negotiation, multi-step technical troubleshooting, calls with financial or regulatory stakes. On live traffic it pays for itself when the voice agent’s resolution path needs multi-step tool calls. Ultra Fast routes the session onto the low-latency fast-model lane, which is what sub-second barge-in needs. Callers interrupt often; if your agent is still finishing a sentence from the previous turn when the caller speaks, your barge-in success rate drops and the line feels broken. Barge-heavy inbound triage lines (order-status lookups, appointment confirmations, routing to a queue) belong on Ultra Fast. Cost Saver is the same Haiku-tier model as Ultra Fast, but off the fast-model lane and tuned for shorter replies — so at very high volumes the per-minute bill drops to roughly a third of Balanced’s. It trades a few hundred milliseconds of first-audio latency for the cheapest per-minute rate. Outbound campaign lines and read-only simple lookups (“what’s my balance?” “when is my appointment?”) belong here.

​
2. Instantiate one, open it in Agents Studio

From the picker, two paths converge on the same draft:
  1. Preview a preset. Click Preview on the card, read the full description and the recommended-for list in the dialog, then click Use preset from there.
  2. Use it directly. Click Use preset on the card — the instantiate POST is guarded so a second click while the first is in flight never mints two agents from one intent.
Either path POST /agents/model-presets/:id/instantiate (owner/admin only) → a draft voice agent is created pre-wired with the preset’s model, transcriber, voice, temperature, and token ceiling. A draft never answers a caller until you activate it. The full API shape is at Model presets on the Agents API page; you can also override the agent name and description in the same call. The picker routes you straight to the new draft agent’s Agents Studio page — the /agents/:id surface — where you rename it, edit the generated starter prompt, and pick a different Cartesia voice. The preset’s job is the model/transcriber bundle; the voice casting and prompt fine-tuning happen in the studio afterward.

​
3. Tune from a preset vs hand-tune provider dropdowns

There is nothing special about an instantiated draft: it is a normal draft agent whose stt_vendor, fast_model lane flag, model, and sampling parameters were filled in by the preset. Everything it ships you can still edit afterward in the agent’s versioned config. The preset is a coordinated starting point, not a lock-in. Use a preset when you are deciding “which reasoning tier should this line run?” — that is a decision, not a configuration. Override by hand (in the studio or via PUT /agents/:id) when you already know the constraint: barge-in lives or dies on the fast_model lane flag, and per-minute cost lives on the model id plus the max-token cap. If you override by hand, set those two knobs together — a cheaper model on the standard lane still won’t respond fast, and the fast lane on a Haiku-tier model still bills at Haiku prices. The preset exists so the three knobs move together coherently.

​
4. Compare two agents before/after with VAQI

Preset choice is not permanent, and the safest re-tier loop runs two agents side by side: keep the current production version on one, instantiate a second from the candidate preset, and read them against each other on the VAQI dashboard (Voice → Voice Agent Quality). Watch the three VAQI families after routing live traffic onto the candidate:
  • Latency — the candidate preset’s real p95 time-to-first-audio vs the estimate on the card. The dashboard flags a warning past 1,200 ms.
  • Barge-in — on Ultra Fast especially, the success rate on interruptions is the number to move; the rate of interruptions itself usually stays.
  • Cost — the estimated per-minute on the card vs the realized per-minute spend, visible per agent.
The full before/after promotion loop (persona simulation, stage-by-stage canary, rollback) is covered in Safely Roll Out an AI Agent; VAQI is the live-traffic read you use inside that loop after a preset lands.

​
Troubleshooting

​
See also