The model is an agent's brain: it sets how the agent reasons, how it sounds under pressure, and what it costs to run. The Builder picks sensible defaults, and you can change any agent's model whenever you like.
Choosing a model
Open an agent's editor drawer → Model & voice (Model & behaviour for text agents). The model picker is a searchable list: type to filter, use the filter chips to narrow by type, and arrow through the matches — each model shows its provider, and a short description appears under the picker when one is available. For voice agents, the Language, Voice and Speech recognition choices cascade from the model you pick, so choose the model first.
You rarely need to agonise here. The Builder chooses models that support everything your design needs, and swapping later is one edit that autosaves into the draft like any other.
Temperature: precise or creative
Below the picker sits a temperature slider running from Precise to Creative. Low settings make the agent consistent and script-faithful — right for confirmations, capture and anything compliance-shaped. Higher settings buy warmth and variety in open conversation, at the cost of predictability. Nothing is saved until you actually move the slider; when in doubt, leave it where the Builder put it and let a test call be the judge.
Agent types differ
Not every model does every job, because agents come in types with different capabilities:
- Phone agents answer real numbers and SIP registrations, speak with a chosen voice, and take keypad input from callers.
- Browser agents take instant calls in a web page — the same technology behind browser test calls.
- Text agents never take calls at all: they chat, and they make ideal subagents — quiet helpers for knowledge, drafting and validation.
Some abilities — transfers, hand-backs, sending keypad tones — belong to particular agent types rather than to models. The Builder steers each design onto a type that supports the plays you asked for, which is one more reason to describe the outcome rather than hand-assemble.
The Workspace rule
One choice is deliberately constrained. Agents holding Google Workspace or Microsoft 365 tools — integration tools or calendar booking — may only run on approved model providers: OpenAI, Anthropic, Google and Ultravox. Providers without a citable commitment not to train on your data are blocked from handling your connected account.
You'll meet the rule in two places: enabling Workspace tools on a team fails with a message naming any agent whose model doesn't qualify, and switching an already-armed agent onto a blocked provider is refused. The fix is always the same — move that agent to an approved model and try again.
Model changes only reach callers when you press Go live, so you can trial a different brain on the draft, test it, and roll back if you preferred the old one.