Create a support agent for your customers

This guide builds an agent your own customers can talk to: it knows only what you gave it, does only what you allowed, and comes to you when it is out of its depth.

An agent you put in front of the public is sandboxed: walled off from everything else you have, knowing only what you put there. That is what makes it safe to hand to a stranger.

Before you start

Write down the ten questions your customers actually ask, and the answers you would give. Then write down what it must never attempt: refunds above a certain value, anything about another customer, anything you want a person to decide.

Decide what it should be able to look up or submit on someone's behalf: an order status, a booking, a callback request. Each becomes one narrow action built for it.

Decide which channel your customers will use. WhatsApp, Telegram and email each need setting up first, and Connect Telegram, WhatsApp, Slack or email covers that. You can add one later without changing the agent. For a chat on your website instead, a sandboxed agent can be opened to anyone with an address: Share an agent has the steps.

Select a concrete project in the header, because a new agent has to belong to one.

Build it

1. Say what you want. Tell your main assistant, in the chat you already use: "I want a support agent for my customers on WhatsApp, answering from our policies and looking up orders." Only the main assistant creates agents, and it starts straight away.
2. Browse first if you prefer. New agent on Manage > Agents (/agents) opens the catalogue beside the chat, grouped by trade, with a Public-facing row marked by a shield; every customer-facing tile carries it. A tier-one support agent and an order status tracker are both there ready made. Click a tile to read what it is for, then say in the chat that you want it. The catalogue only browses; nothing is built until you ask for it.
3. Answer the questions. Your assistant settles the shape in one conversation: what the agent is for, who talks to it, what is out of scope, what it needs to know, where customers reach it. Answer the out-of-scope question properly, because a narrow boundary is most of what keeps the agent safe. When you say go, a new card appears on the gallery with a chat waiting.
4. Give it its knowledge. Ask for a place for it to answer from, and say that customers will read those answers, so the space is created as public. Fill it with your refund policy, your delivery times and the replies you type over and over, and put nothing in it you would not send a customer. Build a knowledge base has the steps and the sensitivity rule. The space appears on Manage > Memories (/manage/memories), connected to your new agent.
5. Build the actions it needs. Say what you want in a sentence: "let it look up an order from the order number and the email address, and give the status and the delivery date". Your assistant builds exactly that, with fixed inputs, one destination, and only the fields you named coming back. Your passwords stay on your side of it, where the agent never sees them. The agent's own page lists every action it is allowed, under Advanced > Tools.
6. Put it in front of customers. Ask your assistant to point the connection you set up at this agent, so messages arriving there go to it rather than to your main assistant. Inside your installation it also has a chat of its own, where you try it.
7. Test it as a customer would. Ask three real questions: one it should answer from its knowledge, one that needs the action you built, and one it has no business answering. The third is the one that matters. Then try to talk it out of its rules, and expect a polite refusal.
8. Decide how it reaches you. Tell it when to stop and fetch a human: an angry customer, a request above your limit, anything it has failed twice. It notifies you, and Admin > Alerts (/admin/alerts) decides which app that lands in, on the Channels tab. Send yourself a test, because an escalation nobody sees is worse than none.

Keeping it safe

Its limits hold whatever anybody types at it. It has no vault, so it holds none of your passwords or keys. It cannot read your files, search the web, see your other conversations or speak to your other agents. It is not on your team, so it is never given a job, and it cannot change any of that, including for itself.

Read what it is being asked. It leaves notes on its Workspace tab: questions it could not answer, gaps in its knowledge, patterns worth knowing. Nothing it writes there returns to its own prompt, which is why a customer cannot teach it anything by talking to it.

It does not learn on its own, so when the notes show a gap, add the answer to its knowledge yourself. Only your main assistant changes a customer-facing agent, and only when you ask.

If something goes wrong

It answers from the wrong material. It reads only what you granted it, so anything odd it quotes is in one of those places: open the space on Manage > Memories (/manage/memories) and fix or delete the note. If it sounds like the template rather than your business, ask in the chat for its role to be rewritten.

It cannot reach you. Check Admin > Alerts (/admin/alerts): the Channels tab needs somewhere to reach you, and Test proves it. On the Routing tab, tick the app you actually watch.

An action fails. The agent says it cannot look something up rather than inventing an answer. Open Advanced > Tools on its page and ask your assistant to test that action. A password behind it may have expired, and replacing it fixes everything using it at once.

Where next

• Build a knowledge base, for filling and maintaining what it answers from.
• Vault and sandboxing, for where your passwords sit while it works.
• Grow your staff, for the rest of the team.
• A support bot on WhatsApp with your knowledge, the same job end to end.
• Share an agent, to open it to the public with a link of its own.