Docs/Quick Start/Create Your First Agent

Create Your First Agent

An agent is your AI chatbot. It has its own name, personality, knowledge base, appearance, and set of channels it's deployed on. You can create as many agents as your plan allows — for example one for customer support, one for sales, and one for a specific product line.

This guide walks through every step of creating and fully configuring a new agent. The initial setup takes about 5 minutes; full configuration typically takes 15–20 minutes.

Step 1 — Create the agent

1

Go to the My Agents dashboard

After logging in you land on the My Agents page. If you have no agents yet you see a blank state with a prompt to create your first one.
2

Click Create New Agent

A creation form appears with two required fields: Name and Instructions.
3

Enter a name

The name is for your internal reference. Choose something descriptive like "Support Bot — ACME" or "Sales Assistant". Visitors see the display name you configure in Chat Interface settings, not this internal name.
4

Write your instructions (system prompt)

This is the single most important configuration field. It defines who the agent is, what it does, and how it behaves. See the Writing instructions section below for a full guide.
5

Click Create Agent

The agent is created and you are taken to its Settings page. From here you configure everything else: the AI model, appearance, knowledge base, and deployment channels.
💡
You can change everything — name, instructions, model, appearance — at any time after creation. Nothing is permanent at this stage.

Writing effective instructions

The instructions field (also called the system prompt) is sent to the AI model before every conversation. It sets the context the model uses to shape every response. Weak instructions produce generic, off-brand replies. Strong instructions produce focused, helpful, on-brand conversations.

What to include

  • Role — Who is this agent? "You are a customer support assistant for Acme Ltd, a UK-based furniture retailer."
  • Purpose — What should it do? "Your job is to answer questions about our products, delivery, and returns policy."
  • Tone — How should it speak? "Be friendly, concise, and professional. Use British English."
  • Boundaries — What should it avoid? "Do not discuss competitor products. Do not provide legal or medical advice."
  • Fallback behaviour — What if it doesn't know? "If you do not know the answer, say so politely and direct the visitor to support@acme.co.uk."
  • Special instructions — "Always ask for the customer's order number before troubleshooting delivery issues."

Example instructions

You are a friendly and professional support assistant for Acme Furniture, a UK furniture retailer.
Your job is to help customers with product questions, delivery timescales, returns, and order status.
Always be warm and concise. Use British English spelling.
If a customer asks about an order, ask for their order number first.
If you don't know the answer to a question, say "I'm not sure about that — please contact our team at support@acme.co.uk for further help."
Do not discuss pricing for commercial or trade orders — direct those customers to our trade team.
💡
Keep instructions to 5–10 sentences. Extremely long system prompts can reduce response quality because the model spends more "attention" on instructions and less on the user's actual question and your knowledge base.
⚠️
Do not put your entire knowledge base or product catalogue in the instructions field. That's what the Knowledge Base is for. Instructions define behaviour; knowledge sources define what the agent knows.

Choosing an AI model

AIChatVault supports multiple large language models. You select the model in Settings → AI Model. Each has different strengths, speeds, and credit costs per message.

ModelBest forSpeedCredit cost
GPT-4oComplex reasoning, nuanced answers, high accuracyFastStandard
GPT-4o MiniHigh-volume support, fast Q&A, cost-sensitive deploymentsVery fastLow
Claude 3.5 SonnetLong-form answers, writing quality, instruction followingFastStandard
Claude 3 HaikuSpeed-critical use cases, simple FAQ botsVery fastLow
Gemini 1.5 ProMultimodal understanding, large context windowsFastStandard
Gemini 1.5 FlashHigh-throughput, budget-friendly deploymentsVery fastLow
ℹ️
You can switch models at any time without re-training. The knowledge base is model-agnostic — all models read from the same indexed sources.

Advanced AI model settings

Inside Settings → AI Model you can tune two parameters beyond model selection:

SettingWhat it doesRecommended value
TemperatureControls randomness. Higher = more creative but less predictable. Lower = more focused and consistent.0.3–0.5 for support bots; 0.7 for creative or conversational agents
Max response tokensLimits the length of each AI reply. Higher = longer answers, more credits used.400–600 for most use cases; 800+ for detailed technical support

Configuring the chat interface

Go to Settings → Chat Interface to control how the widget looks to visitors. These settings affect the web widget, help page, and mobile SDK bubble.

SettingWhat it controls
Display nameThe name shown in the widget header — what visitors see (e.g. "Support Bot" or "ACME Assistant")
Primary colourThe hex colour used for the chat bubble, header, and send button
Widget positionBottom-right or bottom-left corner of the page
Avatar / chat iconUpload a logo or photo; shown in the widget header and on the floating bubble
Welcome messageThe first message the agent sends when a visitor opens the chat
Suggested promptsQuick-reply buttons shown at the start of a conversation (e.g. "Track my order", "Return policy")
Notification bubbleA small card shown above the chat button with a short message to invite visitors to chat
Notification messageThe text shown in the notification bubble (e.g. "Hi! Need help? Chat with us!")
💡
The notification bubble significantly increases chat engagement. Enable it with a short, relevant message like "Got a question? Ask our AI — instant answers 24/7."

Security settings

Go to Settings → Security to control where and how the agent can be used.

  • Allowed domains — List the domains where the widget is permitted to load (e.g. yoursite.com, shop.yoursite.com). Requests from unlisted domains are blocked. Leave blank to allow all domains (useful during development).
  • Rate limiting — Set a maximum number of messages per session or per IP address to prevent abuse.
  • CAPTCHA — Enable to require human verification before the first message. Reduces spam but adds friction.
⚠️
Always add your production domain to the allowed list before launching. Without it, anyone who finds your agent ID could embed your widget on their site and consume your credits.

Lead capture

Go to Settings → Lead Capture to configure the agent to collect visitor contact details during a conversation.

  • Choose which fields to collect: Name, Email, Phone, or custom fields.
  • Set when to ask: before the first message (gate), after a set number of messages, or when the visitor asks for further help.
  • Collected leads appear in the Contacts section of your dashboard.
  • You can export leads as CSV or connect them to your CRM via Webhook or Zapier.

Notifications

Go to Settings → Notifications to receive email alerts about chat activity.

  • New conversation started — get notified when a visitor opens the chat.
  • Lead captured — get notified when a visitor submits their contact details.
  • Human handoff requested — get notified when the agent escalates to a human.
  • Set up to 3 notification email addresses per agent.

Testing your agent before going live

1

Use the Preview chat

Every agent has a built-in preview chat accessible from the top of the Settings page. This lets you test the agent exactly as a visitor would experience it, without consuming your live credits.
2

Ask at least 20 realistic questions

Cover your most common use cases: greetings, product questions, policy questions, edge cases, and things you don't have answers for.
3

Check fallback behaviour

Ask questions outside the agent's knowledge and confirm it responds gracefully rather than hallucinating an answer.
4

Test the lead capture flow

If lead capture is enabled, run through the full flow to confirm fields appear correctly and submissions land in Contacts.
5

Review on mobile

Open the preview on a mobile device or use browser dev tools to check the widget is readable and usable at small screen sizes.
ℹ️
Once you've tested and are satisfied, proceed to the Knowledge Base guide to add your content — then return to Deploy when ready to go live.

Was this page helpful?