AI Agents

Pick who your Agent is in Smart Agent → Agent → Persona. Each persona has its own name, tone and reply language.

• The personas you can choose from depend on your business country and product type. The Replyk team curates and updates them.
• The reply language belongs to the persona, so pick one that speaks the way your customers do.
• Switch personas anytime; try them side by side in the Playground first.

Add your own rules in Smart Agent → Agent → Instructions, on top of everything your Agent already knows about your business.

• Each instruction is a single sentence (max 30 words).
• Use for: brand voice, return policy, delivery promises, upsell rules.
• How many instructions you can add depends on your plan. A meter above the list shows how many you have left.

Scrutiny decides how many subagents check your Agent's work before it goes out. Set it in Smart Agent → Agent → Brain.

• Solo: your Agent alone. Glance: one subagent. Sweep: two subagents. Tribunal: three subagents and a lead.
• There are two separate choices: when it talks (its replies) and when it acts (orders, bookings and anything it sends).
• More scrutiny uses a little more of your balance on every reply it checks.
• Intelligence, in the same place, sets how sharp your Agent is: Smarter uses more of your balance, Cheaper uses less. It appears once it is available for your account.
• Only the business owner can change it.

The Agent is switched on or off per number and per page, not for the whole business. Open the WhatsApp, Instagram or Messenger tab and choose who it answers:

• Everyone: the Agent answers every new conversation.
• Only customers from ads (WhatsApp): the Agent answers people who came from a Click-to-WhatsApp ad, and your team answers everyone else.
• Nobody, my team answers: every conversation goes to your team. Chats the Agent was already handling move to your team when the customer writes again.

The Automatic messages switch pauses everything automatic on that channel, Agent replies included. Incoming messages are still saved and your team can still reply by hand. When you turn it back on, the Agent answers what went unanswered in the last 24 hours.

Your Agent can answer with a voice message instead of text. Choose the mode per WhatsApp number or Instagram account:

• Text only (default): every reply is text.
• When needed: the Agent speaks only while the customer asks to be spoken to or can't read, and goes back to text when they ask for text.
• When the customer sends audio: the Agent speaks whenever the customer's last message was a voice note.
• Always: every reply is spoken. This is the most expensive option, since every reply is billed as text-to-speech.

Settings → Availability & follow-ups controls when the Agent replies and what it does when a customer goes quiet.

• Retry failed replies: if a reply fails, the Agent tries again automatically. If too many conversations fail in a day, retries switch themselves off until an owner or admin turns them back on.
• No-response reminders: nudge a customer who stops replying, after the delay you set.
• To stop the Agent after too many messages, or on anything else a conversation does, use Auto Actions. Auto Actions guide

Test your Agent in a private sandbox before it talks to real customers. Open Smart Agent → Playground.

• Chat in a WhatsApp-style interface.
• See memory, scheduled tasks, and function calls in real time.
• Switch agent profiles to compare behavior.
• Playground replies are billed from the business wallet like any other reply, so it needs a positive balance.

When enabled, your Agent learns from its own finished conversations and gets better at selling your products over time.

• After a conversation ends (won, lost, or cancelled), the Agent reflects on what worked and what didn't.
• Its lessons are combined into a playbook of sales tactics, objection handling and product knowledge that future replies draw on.
• Review what it has learned, and remove anything you disagree with, in Smart Agent → Agent → Self-Learning.
• Toggle it off anytime. The Agent stops learning but keeps using what it already knows.