LLM

Voice agents use LLMs (large language models) that understand the conversation context and respond to the user.

By default an agent runs on the Brosbots-managed LLM - this requires no configuration.

Managed tiers

On the Brosbots (managed) provider you pick a tier, not a concrete model - the same way you do for transcription:

Tier	Model behind it	When
Fast	Gemini 2.5 Flash	What an agent created from a template starts on. The shortest time to the first word: qualifying, routing, confirming a slot, most phone calls.
Smart	GPT-4o mini	A sensible trade-off between pace, price and following instructions.
Accurate	GPT-4.1	Long prompts, branching scripts, tool calls and hard rules the agent must not talk its way around.

The tier is set in the agent editor, on the Role tab. The cost per minute depends on it and is shown in the estimate next to the configuration. The estimate uses the same provider rates as billing and assumes the top of the range: four model requests a minute, each with the full prompt, and an agent speaking for the whole minute.

The model is usually the largest part of the pause before the agent answers. If an agent feels sluggish, start with the Fast tier and only then touch turn detection. Move to Accurate when the agent loses its instructions on a faster tier.

From the Pro plan up you can use your own provider accounts (OpenAI, Google, Groq): add a key in Account → Provider keys, then pick the provider and model in the agent configuration.

When the model does not answer

A single model failure (a 5xx, a rate limit, a dropped stream) no longer means silence on the line:

1. A request that failed, or produced neither a word nor a tool call within 10 s, is sent once more.
2. When that fails too, a fallback tier on another provider answers: Fast (Gemini) stands in for Smart and Accurate (OpenAI), Smart stands in for Fast. For the next 60 s the call starts with the fallback, so every turn does not wait for two failures first.
3. When nothing works, the agent says a short apology (“I’m sorry, we’re having a technical problem…”) and ends the call with the service_failure disposition.

A reply that has already started being spoken, or has called a tool, is not repeated - a retry would say the same words twice or book the appointment twice.

Set the fallback with workflow_configurations.llm_fallback_tier (fast, smart, accurate or none). An agent on its own provider key has no fallback until you set one - the fallback runs on the Brosbots key.