Skip to main content

​
What is this?

Open an agent from the Meta Business Agent page to configure it. Its tabs are Overview, Behavior, Business info, Knowledge, Skills, Tools and Test (see Test). This page covers the setup tabs.

​
What can I do here?

  • Choose who gets AI replies and what happens at handoff
  • Cap usage so cost stays predictable
  • Tell the agent about your business, and give it FAQs, websites and files to answer from
  • Write skills — and rich messages like buttons and Flows
  • Connect your own systems as tools
  • Start from a ready-made template

​
How to use it

​
Behavior

Who gets AI replies, hand-off to your team and follow-ups.
  • Let the agent reply — the rollout switch (the same as Agentic Mode). Choose Everyone, or Allowlisted numbers only and add numbers in international format (for example +15551234567) in the Allowlist. Use the allowlist to try the agent on your own number first.
  • Message before handing over — Standard message (a fixed message), Agent writes one (a fitting message for each chat), or My own message.
  • Never say — one phrase per line. If a reply would contain one, it isn’t sent.
  • Usage limits — cap the agent so cost stays predictable. A limit is Tokens (across your business) or AI replies (per conversation), Per day, 7, 14 or 30 days. When a limit is reached the agent finishes its current reply, stops answering new messages and hands the chat to your team; it starts again when the period rolls over. Each type and period can only be used once.

​
Business info

What the agent says about your business. Edit and save it; Reset clears your changes.

​
Knowledge

Questions, websites and files the agent answers from. Click Add source and choose FAQ, Website (the number of pages crawled is shown) or File. Edit or remove a source any time.

​
Skills

Instructions for what the agent does in specific situations. Click Add skill.
  • Name — lowercase English letters, numbers and hyphens, like product-return-policy. Only the name needs English.
  • When should the agent use it? and Instructions — in any language. Write clear steps, and don’t give two skills the same trigger.
  • A skill shows its status: Active, Being reviewed, or Blocked (Meta blocked it, usually because it asks for personal details — edit it to clear the block).
Rich messages are skills that send something other than text: a Link button, Reply buttons (up to three), a List menu, an Image, a Location, a request for the customer’s location, a Link carousel, a Reply carousel, or a WhatsApp Flow. Describe when to send it and include everything the message needs (text, button label, link). Only published Flows can be switched on.

​
Tools

Outside services the agent can use. Click Add connector and fill in the service’s name, what it is for, its Service address, and how it signs in (No sign-in, API key, or Client ID and secret (OAuth)). Then Add tool for each action — a name like check_order_status, when the agent should use it, the request (action and path) and the details the agent needs to ask the customer for. Mark Customer must sign in first when the service needs the customer’s own login. Make sure the service answers before you switch the agent on.

​
Templates

Click Use a template to start from a ready-made setup — for example Customer support, Order status, Sell a single item, Product recommendations, Stock by location, Basket over several messages, Sign in with a one-time code, Bookings and rentals, Changes and refunds after booking, Several parts one purchase, or Lead generation. Review what will be added (instructions, FAQs, rich messages, tools); nothing you already have with the same name is changed.

​
Troubleshooting / Technical Notes

  • “Auth config must be valid JSON” / “Request definition must be valid JSON”. You’re editing a tool as JSON — fix the syntax, or go Back to the form.
  • “This Flow is not published yet.” Publish the Flow first, or save the rich message as off.
  • A skill is Blocked. Meta blocked it — usually because it asks for personal details. Edit it.
  • “Each type and period can only be used once.” Edit the existing usage limit instead of adding a duplicate.
  • The agent doesn’t answer. Check Agentic Mode, the audience and allowlist, and that no usage limit has been reached.
Last modified on October 29, 2026