Before you start
Agents is the base of AssistantLabs: the AI agent that answers your customers on WhatsApp, Instagram, Messenger, LinkedIn, email and your website, in your tone and their language, with one inbox behind all of them. It learns from your site, documents, interview, catalog and past conversations, follows your rules and conversation scripts, tags and routes chats, reaches out first with follow-ups, and hands over to you when a person is needed. Every workspace has it. There are two ways to set it up, and you can switch between them at any step. They change the same agent.
- An AssistantLabs account. It is free to start at assistantlabs.io/app.
- Your website address. No site? Teach it from a document or a short interview instead.
- For the Claude route: Claude connected to your workspace. Install the plugin once, as in Connect to Claude. ChatGPT works too: Connect to ChatGPT.
- Later, for a channel: your WhatsApp Business number and Meta business account. You do not need them to build and test.
Use the switch above each step. With Claude gives you a prompt to paste and what it will do. In the app shows the screens and where to click; Logos, the AI builder inside the app, can do the same job from a chat (Meet Logos).
Create it from your website
Your website already says most of what customers ask: what you sell, what it costs, when you are open. Reading it is the fastest way to give the agent a first version of all of that.
Read my website https://willowandco.com and build my AssistantLabs agent from it. Use the business’s own name and its own words. Before you save anything, show me what you understood: hours, prices, delivery, returns, and anything the site contradicts itself on.
What Claude does, in order:
read_websitereads your entry page and the pages that matter (about, contact, prices) and returns the text and the phone, email and address found in links.create_agentmakes the agent, named the way your site names the business. It is skipped if you already have one.add_agent_sourcerecords the site under Settings → Sources, so you can see where each fact came from and remove it as a set.- After you confirm,
patch_agent_modulefills the business details, FAQ, catalog, links and guidelines, andpatch_agent_languagesets the language.
Prefer a command? Type/al-website. It runs the same read. No website? Use/al-documentfor a price list or/al-interviewto be asked about your business one question at a time.
New to the app? On Home, describe what your business does in a sentence, paste your website link, or start from a template for your type of business. Logos builds a first version of the agent from it.
The same scan lives in Settings: Build → Sources, then Create Agent from Website.
Make it sound like you
A new agent answers correctly and sounds like nobody. Its greeting, tone, how long its replies run, emoji and grammatical gender (which matters in Hebrew) are set separately from what it knows. They are the part only you can answer.
Here is how I would really answer a customer who asks “is there space tomorrow morning?”: “Hi! Yes, I have 10:00 and 11:30. Which one suits you?” Set my agent’s greeting, tone and reply length from that. I am a one-person business, so it speaks as I, female.
- Claude calls
set_agent_personawith the greeting, tone, gender, response length and emoji. - It is a different place from the knowledge modules, which is why a plain module write is refused for the persona.
- Check it afterwards: “show me my agent’s persona” reads it back with
get_agent_config.
Field by field: The Persona guide walks every field, including the languages it answers in.
Check what it knows
Correcting a wrong opening hour is faster than writing twenty answers from nothing. Look at what the agent has, then fill the gaps with the answers customers will actually ask for.
Show me everything my agent knows about my business, and list the three questions customers will certainly ask that it cannot answer yet.
Add this to the FAQ: “Do you ship to Eilat?” → “Yes, 3–5 business days.” Show me exactly what you changed.
From now on, always mention the 14-day return window when someone asks about returns.
get_agent_knowledgereturns a readable summary of the business details, persona, FAQ, links, products and rules. With conversation history,conversation_statsshows what people ask against what it knows.patch_agent_moduleadds the answer to the FAQ module. It takes the whole module, so Claude reads it first and shows you the diff.set_extra_instructionssets a free-text rule on top of everything else. It replaces the previous overlay in full, so ask Claude to show you the current one first.
Open Build → FAQ and add a question. Or tell Logos and it saves the answer for you.
Products and prices belong in the Catalog; how to add a question is in the FAQ guide.
Try it before anyone else does
What you see in the Test panel is exactly how the agent will answer real customers. Ask it what they ask, including the awkward ones.
Write me five questions a real customer of mine would ask, including one awkward one, and tell me where to try them.
generate_test_customer_messagewrites the customer’s side, based on how your agent is configured. Add a scenario (“an angry customer whose parcel is late”) to steer it.- Claude sends each question to the agent and shows you the replies. You can also paste them into the Test panel, or open the agent’s chat link, and read the answers there.
When an answer is wrong: Say what it should have said and Claude fixes the module, or files it for you to review with propose_training. See Training.If an answer is off, click Edit response under it. Your correction becomes a training example on the spot.
Connect a channel, and leave it muted
Connect WhatsApp, but keep Auto-reply off. The agent is connected and listening, and real customers’ questions arrive without it answering yet.
Which channels is my agent connected to, and what is still missing before customers can reach it?
list_channels. Leave Auto-reply off for now, as shown in the next tab.Step by step, with Meta’s own screens: Connect WhatsApp or Connect Instagram.
Watch real chats, then turn it on
Give it two or three days. Real customers ask things you would never think of at your desk, and with Auto-reply off you read them without anything going out.
Read the conversations that came in over the last three days. Which questions would my agent have answered wrong or not at all, and what should I add to fix each one?
list_conversationsandget_conversationread what came in;conversation_statscounts topics and busy hours.- Say “add those” and it writes the answers with
patch_agent_module. Hand-picked wording stays yours.
When the answers look right, tell Claude “turn Auto-reply on for WhatsApp”, or do it yourself in Build → Channels by switching Auto-reply on.
Before you flip it: Set a notification so you are pinged when a chat needs a person, and a few tags to keep complaints and orders from getting buried.
Rules, tags, follow-ups and alerts
Answering well is the start. The rest of what the agent does lives in four more places under Build, and each can be set from Claude or from the app.
Add two rules to my agent: never promise a delivery date, and always offer to pass the chat to me when someone asks for a refund.
Set up a tag called “Complaint” that the agent puts on any chat where the customer is unhappy.
Build me a scenario for booking a first appointment: ask what service, then the day, then take a name and phone number.
Which follow-ups are waiting, and draft one for the customer who asked about the linen dress and went quiet.
Ping me on WhatsApp whenever a chat needs a person.
- Rules:
patch_agent_moduleon the guidelines module, one rule per item. - Tags: the labels module, written in plain words, so the agent decides which chats get which tag.
- Follow-ups:
list_follow_ups,draft_follow_upandupdate_follow_up. A draft is for you to review. - Alerts:
setup_owner_notificationsandcheck_owner_notifications.
Under Build, the side menu has Guidelines (the rules it must follow), Scenarios (a guided flow with steps), Follow-ups, Tags and Notifications.
- Scenarios: guided, branching conversations such as taking a booking or qualifying a lead. Ask Claude to build one, or edit it in the app. Scenarios guide.
- Follow-ups: the agent reaches out first to a customer who went quiet. Follow-ups guide.
- Tags and notifications: tag chats on their own and ping the right person. Tags, Notifications.
- Tools it can use: your catalog and Google Sheets, Shopify, WooCommerce and Wix stores, Calendly and Google Calendar, Monday and Fireberry, plus your own HTTP API as a custom integration (
create_custom_integration). Integration map. - Beyond one agent: several agents in one workspace, team members, and links to Booking, Sales, Helpdesk, CRM, Marketing and Tasks. See the Agents app guide.
Improve it every week
An agent that gets one review a week is a different business in six months. Find where it answered badly, fix it, and let the fix apply to every chat after it.
Go through this week’s conversations and find where my agent answered badly. For each one, file a training suggestion for me to review. Do not change anything live.
- Claude reads the week with
list_conversationsandget_conversation. - Each problem becomes a suggestion with
propose_training. It sits as pending, and nothing changes until you approve it. - Approve in the Training tab, or say “approve the first two” and Claude uses
approve_training_proposal.
- Once a week, review the Training tab, or run the prompt above.
- Keep an eye on the questions the agent has no answer for. Those are worth more than a polished tone.
- Want the slower, week-by-week version of all of this? See Get Your Agent Live in 3 Weeks, and the Agents app guide for everything else the app does.