Agent Creation Guide

Agents are the core of the platform. Each agent handles inbound or outbound calls with its own voice, persona, and conversation logic.

---

Two Modes

	Freeform	Flow
Structure	Single system prompt	Multiple nodes, each with its own prompt
Tools	Available throughout the call	Per-node
Control	LLM decides everything	Structured, step-by-step
Best for	Open-ended conversations	Scripted, multi-stage conversations

Use freeform when: The caller can ask anything at any time — support, FAQ, lookup.

Use flows when: The conversation has stages — qualify → book → confirm, or greet → verify → offer → close.

---

Contents

Doc	What it covers
Freeform Agents	Simple agents, multilingual, inline tools
Flow Agents	Multi-node flows, branching, escape hatches
Tools	Webhook tools, pre-call tools, pre-actions
Outbound Calls	Context variables, outbound dialling
Objection Handling	Call me later, not interested, transfer to human, DNC
JSON Import & AI Creation	Full JSON schema, AI prompt template, common mistakes

---

Quick Start

Minimal freeform agent (API):

curl -X POST https://your-domain.com/api/agents \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "support-agent",
    "prompt": "You are Maya, a friendly support agent for Acme Corp. Help customers with order queries.",
    "greeting": "Hi, this is Maya from Acme. How can I help you today?"
  }'

Minimal freeform agent (import JSON):

{
  "version": "1",
  "agent": {
    "name": "support-agent",
    "prompt": "You are Maya, a friendly support agent for Acme Corp.",
    "greeting": "Hi, this is Maya from Acme. How can I help you today?"
  },
  "tools": [],
  "flow_nodes": []
}

Upload via Dashboard → Agents → ↑ Import, or POST /api/agents/import.

---

Pipeline at a Glance

Every call goes through:

Audio In → VAD → STT → LLM → TTS → Audio Out

• VAD — detects when the caller starts and stops speaking
• STT — converts speech to text (Deepgram, Sarvam, OpenAI, ElevenLabs)
• LLM — generates the response (OpenAI, Google, Grok)
• TTS — converts text to speech (Cartesia, ElevenLabs, Sarvam, Deepgram, OpenAI)

Default stack: deepgram nova-3-general + gpt-4.1-nano + cartesia sonic-3.

---

Auto-Injected Rules

The platform appends these to every agent's system prompt automatically — you do not need to add them:

• Max 2 sentences per response (unless the caller asks for detail)
• No markdown, bullet points, or special characters
• Spell out numbers in full
• No filler openers ("Certainly!", "Of course!", etc.)
• Call end_call when the conversation is complete

---

Auto-Injected Call Metadata

The LLM always receives:

Call info: This is an inbound call. Caller phone: +919876543210.

For outbound calls with context variables:

Call context:
- customer_name: Ravi Kumar
- invoice_amount: 4500