Agents

An Agent is one voice AI assistant with its own personality, voice, knowledge, and abilities — think of it as a single "employee" with a specific role, like a support rep, an appointment scheduler, or a salesperson. Each agent carries its own system prompt, its own provider setup (the model, the ears, and the voice), an optional knowledge base, and optional tools it can use mid-call.

Everything on this page happens in the dashboard at your-infravoice-host — no code required. Your workspace even ships with three ready-made demo agents so you're never starting from a blank screen.

New here? A great first goal: create an agent, write a clear system prompt, add your provider keys, publish it, then open Voice Test and have a real conversation with it — about 20 minutes.

What makes up an agent

Part	What it is
System prompt	Plain-English instructions that define who the agent is and how it behaves. The single most important setting.
Providers	The three engines every agent needs: an LLM (the brain), Speech-to-Text / ASR (the ears), and Text-to-Speech / TTS (the voice).
Knowledge Base	Your own facts — FAQs, price lists, policies — so the agent answers accurately instead of guessing.
Tools	Actions the agent can take mid-call, like looking up an order or booking a slot.
Handbook	Reusable prompt snippets (shared policies, tone, disclaimers) you can attach to many agents.
Checklist	The points an agent should cover on every call, ticked off live by a team leader's monitor.

Your agents

The Agents page (in the left sidebar under Build) lists every voice agent you've built.

Click any row to open and edit an agent.
Use Create an Agent to build a new one.
The Active badge means the agent is published and ready to take calls.

Create & configure an agent

Click Create an Agent (or open an existing one) to reach the editor. Across the top you'll find the tabs that let you fine-tune everything:

Configuration · Providers · Tools · Knowledge Base · Handbook · Checklist · Telephony · Flow Builder · Tests · Analysis · History

On the Configuration tab, fill in:

Agent Name — something memorable, e.g. "Front-desk Booking Bot".
Description — a short note for your own reference (optional).
System Prompt — the most important field. Describe who the agent is and how it should behave, in plain English.
Welcome Message — the first thing the agent speaks when a call connects. Leave it empty for no greeting.
Language and Connection Type — pick your primary spoken language; leave the connection type on its default if unsure.

Click Save Changes when done.

Writing a great system prompt

The system prompt defines the agent's personality, knowledge, and boundaries. A well-structured one usually covers:

Identity — "You are Emma, a warm AI nurse assistant for City Clinic…"
Behaviour — tone, response length, conversation style.
Domain knowledge — key facts, products, policies.
Boundaries — what the agent should not do.
Edge cases — how to escalate, transfer, or handle unknown questions.

Example

You are Emma, a warm AI nurse assistant for City Clinic.

Help patients book, reschedule, or cancel appointments.

Keep replies to 1–3 short sentences and use the patient's name when you have it.

Always confirm the date and time before finishing.

Never give medical advice — offer to connect the caller to a nurse instead.

Tip: Use the Knowledge Base for facts ("our hours are 9–5") and the system prompt / Handbook for behaviour ("always greet by name, never argue with a customer").

Providers — give it a brain, ears, and a voice

An agent needs three engines before it can take a live call. Set them on the Providers tab:

Speech-to-Text (ASR) — how the agent hears the caller (e.g. Deepgram).
LLM (the brain) — the model that decides what to say (e.g. OpenAI, Gemini, Anthropic).
Text-to-Speech (TTS) — the voice the agent speaks with (e.g. ElevenLabs, Cartesia).

Save each one with Save LLM, Save ASR, and Save TTS. You can also set an optional Analysis LLM used for post-call summaries and QA — leave it empty and the agent's main LLM is used instead.

All three (LLM + STT + TTS) are required before an agent can take a live call. You bring your own provider keys — InfraVoice hosts no models.

Save keys once, reuse everywhere

Instead of pasting a key into every agent, save it once under Build → Credentials (Provider Credentials) and reuse it across every agent:

Click Add credential.
Choose the Type — LLM, Speech-to-Text, or Text-to-Speech.
Pick the Provider, give it a Name, and paste the API Key (plus an optional Base URL, Model, or Voice ID).
Save. The key is validated before it's stored, encrypted at rest, scoped to your workspace, and never shown in full again.

Knowledge Base — give it the facts

Feed your agent your own content so it answers accurately. Manage knowledge globally under Build → Knowledge Base (Add Memory), then attach a specific entry to an agent from that agent's Knowledge Base tab.

Handbook — reusable prompt snippets

Under Build → Handbook you can write reusable prompt snippets — shared policies, tone, or disclaimers you want to keep consistent across a fleet of agents. Create one with New snippet, then attach it from an agent's Handbook tab.

Tools — let the agent do things

Tools let your agent take actions mid-call — look up an order, check a calendar, book a slot — instead of only talking. The LLM decides when to use a tool based on the conversation, and the result is folded back into a natural spoken reply.

Add a tool on the Tools tab: give it a clear name and description, and (for tools that call your own service) an endpoint. The agent uses it when it makes sense.

Connectors: Under Build → Integrations you can connect third-party services (e.g. a calendar) once, then enable them per-agent on the Tools tab.

Checklist (optional)

The Checklist tab lets you list the points an agent should cover on every call (e.g. "verify identity", "confirm the appointment time"). During live calls, a team leader's Live Monitor ticks these off automatically as they're met — a great quality signal.

Flow Builder — step-by-step conversations (optional)

Most agents work great with just a system prompt. But when a call must follow exact steps — collect specific information in order, branch on the answer, or call a tool at just the right moment — the Flow Builder tab lets you draw the conversation as a diagram.

Drag blocks from the palette and connect them: Welcome · Conversation · Question / Decision · Tool Call / Variable · Transfer · End. The graph saves automatically, but it only goes live when you click Activate Flow Mode — until then, calls keep using the agent's single system prompt.

The Flow Builder is best used on a desktop; on a small screen you'll see a read-only summary of the flow instead of the canvas.

Publish — make your changes go live

Saving edits creates a draft. Calls keep using the last published version until you publish the new one. Use the version banner at the top of the editor:

A brand-new agent shows Not published yet → click Publish v1.
After edits, the banner shows a draft → click Publish v2 (v3, …) to push it live.
When everything is published, the banner reads Live.

You can review earlier versions and roll back from the History tab.

Test your agent

Once an agent is published with all three providers set, open Voice Test (the Voice Test Studio) from the sidebar, pick your agent, and click Start Call — a real, live voice conversation right in your browser, no phone needed. Watch the live transcript, use Mute if needed, and End Call when you're done.

For hands-off checks after a prompt change, use Build → Simulation Tests (or the agent's own Tests tab) to run LLM-driven test conversations and review each result.

Next steps: To put your agent on a real phone number, see the Telephony tab and the User Guide's Twilio setup. To hand live calls to human agents, set up a Call Center project.