Agents

An agent is the AI persona that runs each conversation. Configure the script, language, voice, models, call behavior, analytics, and inbound phone assignment from one builder.

Create an agent

1. Open Agent Setup from the sidebar.
2. Click New agent. You are asked How do you want to start? — pick Auto Build to describe the call and have it written for you, or Start from Template to begin from a ready-made agent (or the blank Custom template).
3. Rename the agent from the summary header so it is easy to select later in campaigns and batches.
4. Fill the Agent, LLM, Audio, Engine, Call, Analytics, Inbound, and Tools tabs.
5. Click Save from the header actions before using the agent at scale.

Build with AI (Auto Build)

Rather than writing everything by hand, choose Auto Build in the New agent chooser and describe the screening call in plain English. The system generates a complete agent — welcome message, full call prompt, and a matching rating template — that you review, tweak, and save.

1. Give the agent a title and pick its spoken name (e.g. Isha) — the name the agent says on the call.
2. In Describe the screening call, type what you need (role, skills, number of questions, what to evaluate). Not sure what to write? Click Examples (top-right) to drop in a ready-made description.
3. Optionally fill the structured fields: Calling from (the company the agent names in the opener — defaults to Veytrix AI), Role, Client / hiring company, Number of questions, and Focus areas. All optional — they sharpen the result.
4. Click Generate, review the welcome / prompt / rating template, then Save this agent.
5. Finish by attaching a phone number and a voice.

Agent tab

This tab controls the conversation purpose.

• Agent type - choose Real conversation for two-way AI calls, or Announcement for one-way calls that play the welcome message and hang up.
• Welcome message - the first line spoken to the caller. Write it as fixed text; voice agents don't substitute CSV columns, so don't use {name}-style placeholders (they are read aloud literally).
• Agent prompt - the core instruction set: goal, tone, required questions, disqualification rules, and closing behavior.

You are a screening agent for a Java Full Stack role.
Greet the candidate by name.
Confirm availability for the role and location.
Ask about Java, Spring Boot, React, SQL, and project experience.
Ask one follow-up when an answer is unclear.
Keep each response under two sentences.
If the candidate asks a policy question, answer only from the linked knowledge base.
End politely after all required questions are complete.

The conversation pipeline

Veytrix streams call audio through three pluggable providers, each chosen per agent:

LLM tab

• Provider and model - choose OpenAI, Groq, Cerebras, or Gemini, then select a supported model.
• API credential - select the workspace credential from Settings > AI Services. Hidden on managed workspaces, which use our keys — see step 2 of Getting started.
• Tokens - maximum generated tokens for each response. Higher values allow longer replies but can slow the call.
• Temperature - lower values are more consistent; higher values are more varied.
• Knowledge bases - multi-select reference documents that the agent can use during the call.

Audio tab

• Language - the language the agent speaks. Eleven are available: English, Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi, and Odia.
• Also understand and speak - extra languages the caller and agent may use. Sarvam detects each utterance and the reply is spoken in that language.
• Speech-to-text - choose Sarvam, Deepgram, or Cartesia and select the transcription model.
• Keywords - add important names, skills, products, or acronyms to improve recognition.
• Text-to-speech - choose Cartesia, Sarvam, or ElevenLabs, then pick a voice from the searchable voice picker, plus model, speed, and credential.
• Voice quality - for Sarvam, choose telephony, standard, or high-quality sample rate based on the model.

Conversation engines

• Pipeline (default) — the modular STT → LLM → TTS chain above. Maximum control; ~0.8–2s response.
• Gemini-Live — Google's speech-to-speech model; lowest latency, fewer knobs.

Engine tab

• Pipeline uses separate STT, LLM, and TTS settings. Use it when you need provider control, knowledge bases, and detailed tuning.
• Gemini Live uses one speech-to-speech model. Use it when low latency is more important than separate provider control.
• Interrupt after words controls how quickly the agent can interrupt after the caller starts speaking.
• Endpointing controls how quickly the system treats caller speech as complete.
• Response rate lets you trade speed against smoother delivery.

Call tab

Use this tab to control telephony and call safety behavior.

Analytics tab

• Call rating template decides which rubric and columns are used on the Ratings page.
• Analytics webhook URL sends execution data to your own endpoint.
• Summarization creates a post-call summary.
• Extraction captures custom structured fields from the transcript.

Inbound tab

Assign an inbound-capable number to the agent. One inbound number can ring exactly one agent, and numbers already assigned to other agents are hidden. Manage the full number list from My Numbers.

Tools tab — Google Calendar

Turn on Google Calendar to auto-create a calendar event (with a Google Meet link) shortly after a call is rated — using the date and time the candidate confirmed during the call. The event is created on a host’s Workspace calendar and the candidate is invited.

Key settings

Try the agent before you dial anyone

Three ways to hear or read an agent, all from the header actions in Agent Setup:

Test and publish

1. Save the agent.
2. Start a Web Call for a fast, free pass over the whole conversation.
3. Then use Get a call from agent with your own phone number, to hear it over a real line.
4. Check greeting, pronunciation, language, interruption timing, and final message.
5. Open Call History to inspect transcript and summary.
6. Open Ratings to confirm the selected rating template produces the expected columns.
7. After validation, select this agent when creating a campaign or batch.