Agents

An agent is a configured AI voice that handles calls on your behalf. Each agent has a prompt, a voice, an LLM, and an optional set of tools. This page covers every field on the agent editor and when to change it.

The system prompt

The prompt is the single biggest lever on agent behaviour. Keep it short, direct, and grounded in concrete examples. A strong prompt has four parts:

  1. Role — who the agent is and who it serves.
  2. Goal — what a successful call looks like.
  3. Constraints — what it must never do or say.
  4. Style — tone, response length, and language.
Rule of thumb: if your prompt is longer than 400 words, you are probably over-specifying. Move edge-case behaviour into tools or the knowledge base instead.

Voices

Voices come from our upstream TTS providers. Each has a preview button in the editor. When picking one, match:

  • Accent to your caller base (US, UK, neutral).
  • Energy to the use case (calm for support, brighter for outbound sales).
  • Gender and age to your brand.

Voice choice affects per-minute cost. Premium voices add roughly 2–4 cents per minute on top of the base rate.

LLM selection

The default model balances latency, quality and cost for most voice use cases. Switch models when:

  • Responses feel sluggish — try a smaller, faster model.
  • The agent hallucinates or misses nuances — move up one tier.
  • You need function-calling reliability for tools — use a tool-tuned model.

Tools

Tools let the agent call your backend mid-call. Common examples:

  • check_availability(date)— book appointments.
  • lookup_order(order_id)— fetch customer status.
  • transfer_to_human(reason)— escalate to a live agent.

Each tool is an HTTPS endpoint on your server. Xenith posts a JSON payload with the arguments the LLM extracted, your endpoint returns JSON, and the agent speaks the result. Tool endpoints should respond within 2 seconds to avoid awkward silences.

Guardrails

Guardrails are post-generation checks that can interrupt the agent. Use them to:

  • Block profanity or competitor mentions.
  • Enforce a hard maximum call duration.
  • Refuse to discuss regulated topics (medical, legal, financial).

Testing before you deploy

Every change to a live agent is saved as a new version. Before flipping a new version to your phone number, test it with the browser simulator. Run at least five scripted scenarios and one free-form call.

Next