Meta Business Agent can now return the rich WhatsApp messages it would send during Agent Test. You can inspect buttons, lists, carousels, documents, catalog items and Flows before a real customer receives them.
The response is still a simulation. It does not prove media access, device rendering or completed action. Inspect the component, then prove the operating path in an allowlisted conversation.
Understand what changed
Meta added a ui_components array to Agent Test on 12 November 2026. The next day it documented more interactive types, including a carousel with two quick replies per card, document messages, multi-product lists and catalog messages.
Each component identifies its type, source and fields. Buttons expose labels and URLs, lists expose rows, and documents expose a URL, filename and caption. Product messages expose catalog identifiers.
This extends the general Meta Business Agent launch test. It does not replace the broader release decision.
Freeze one customer task
Choose one clear task, such as: “Show three waterproof jackets under my budget, then let me ask for help.” Record the component, products, fallback and human destination.

- 1PromptUse one bounded customer task.
- 2ComponentConfirm type and source.
- 3FieldsCheck labels, URLs and product IDs.
- 4DeviceProve rendering and media access.
- 5HandoffKeep context and a visible owner.
- 6ReceiptStore evidence and the release decision.
Do not test every type together. Fluent text can hide a wrong product set or unusable action. Use the guide to WhatsApp buttons, lists and Flows to choose the lightest format.
Read the response as release evidence
Inspect the text response and the ui_components array separately. A good sentence does not prove that the correct component was built.
| Evidence | Check in Agent Test | Block release when |
|---|---|---|
| Type and source | Expected component and configured or built-in source | The agent chooses the wrong interaction |
| Customer wording | Button labels, list titles, captions and filenames | A label is vague or promises the wrong result |
| Destination | URLs, Flow identifiers and product identifiers | A target is missing, stale or unexpected |
| Order and scope | Component sequence, rows, cards and products | The customer sees too much or the order is confusing |
| Stop state | Handoff or no-response reason | No accountable next owner exists |
Meta says failed or gated messages are omitted from ui_components. An empty array needs investigation; it does not prove the agent deliberately chose plain text.
Test the fields that can mislead a customer
For buttons and lists, read every label as a customer. “Continue” is weak if the tap starts payment or confirms an order. Verify that actions map to the intended step.
For carousels and product lists, check identifiers against the current catalog. The agent should not present an unavailable item as confirmed stock. The authoritative commerce system must supply live inventory state.
For documents, verify the filename and caption. Treat the HTTPS URL as untested until a safe check fetches it, and confirm that media exposes no customer data. Apply the AI customer reply review to prices, dates and promises.
Move to a real controlled conversation
Meta explicitly says Agent Test does not fetch media URLs or confirm production delivery and device rendering. After the simulated fields pass, enable the agent only for an approved allowlisted consumer and repeat the same journey on a real phone.
Check card order, text truncation, taps, links, documents and Flow completion. Try an expired link, missing product and unexpected reply. A safe result may be a clear owned stop.
This is the step where the AI release gate matters. A simulation is component evidence; an allowlisted conversation is operating evidence.
Prove the reply and human path
Tap every allowed action. Confirm the reply reaches the expected workflow and record. Ask for a person midway through the carousel or Flow. The agent should stop, preserve context and expose a handoff reason.
The team needs a visible owner, recent messages, the customer’s selection and next action. If a configured DripTell inbox is involved, verify the supported connection and handoff. Do not infer a native Meta Business Agent integration from a generic inbox capability. The eligibility and ownership map explains the boundary.
Keep a release receipt
Store the configuration version, phone number, prompt, conversation identifier, expected component, returned fields, failures, real-device evidence, owner and decision. Repeat critical cases after an instruction, catalog, knowledge, connector or Flow changes.
Inspect what the agent intended to send, then prove what customer and team experience. Neither layer is enough alone.
Frequently Asked Questions
Does Agent Test prove an interactive message will be delivered
No. It proves which supported rich message the simulated sender accepted. Meta says it does not confirm media accessibility, production delivery or rendering on a particular device.
What should be checked first
Confirm the component type and business destination before polishing wording. A beautiful carousel with the wrong products or URL is still a release failure.
Should every component be tested on a real phone
Test every component used by the journey, plus its main failure and handoff paths. Do not test unused types merely because the API supports them.
Sources
Want help applying this guide?Ask the DripTell team
From reading to action
Want help applying this to your business?
Share your details. We will read the article you are on and start with a useful answer for your team.
DripTell Editorial
Practical guidance reviewed by the DripTell product and customer workflow team.
See how DripTell checks product claims, uses primary sources and handles corrections.
Editorial and source policy



