Skip to main content
Three ways to see what an agent does before it talks to everybody: the test drawer in the studio, test mode on your own number, and the draft button in the inbox.

The test drawer

Test agent in the studio header opens a chat panel. It is reachable from every tab, so you can change a ground rule and try it in the same breath. It runs the same assembly the live agent runs — the same instructions, the same knowledge retrieval, the same tools. A test surface that differed from production would teach you to trust behaviour you will not get. Under each reply, a collapsed line shows what the answer was built from:
  • sources — the knowledge excerpts retrieved, with the document title and the passage
  • tool calls — each tool the agent ran, its arguments, whether it succeeded, how long it took, and the result it was given
If the agent decided to escalate, the bubble says Would hand off to a human here. Three things differ from a live conversation:
  • No customer is attached. Tools that need one customer — order lookup, the payment link tool, automation triggers — say so instead of reading or writing somebody’s real records.
  • Nothing is sent to anyone. No message goes out and no conversation is touched. Send approved templates is never offered here, because there is no WhatsApp conversation to send one in.
  • No channel is known, so replies follow the plain-text formatting rules used for Instagram and the website chat, without WhatsApp’s *bold* markup. See How replies are formatted.
Tools that write still write. A Google calendar booking, a spreadsheet row, a Cal.com booking, a deal or one of your custom actions called from the test drawer does the real thing. A Cal.com booking or a deal made here has no contact attached, because there is none. Only sending a message is suppressed.
The drawer is metered exactly like production — an unmetered test surface running on our key would be an open AI endpoint with a login page in front of it. On built-in AI, testing spends credits; on your own provider key nothing is metered and your provider bills the calls.

Test mode

On the Behaviour tab, Try it on your own number first:
1

Add your test numbers

Up to three, with the country code — +91 98765 43210. These are matched exactly against the contact’s phone number.
2

Switch Test mode on

The switch is disabled until at least one number is listed; with none, the agent would answer nobody. Clearing the last number turns test mode off again.
3

Message the business from that number

The agent replies to you normally. Every other customer is left for your team, exactly as if the agent were off for them.
Test mode is a phone-number allowlist, so it silences the agent on Instagram DMs and the website chat too — those conversations have no phone number to match. That is the correct reading of “only reply to my test numbers”, but it does mean test mode is not a way to trial an Instagram agent. The agent’s row and the studio header both show a Test mode pill while it is on, so it cannot be left on by accident.
On built-in AI credits there is no separate message quota for test mode — the wallet is the cap. On your own provider key nothing is capped at all: your provider bills you directly, and a limit we invented would be theatre.

Drafting in the inbox

The sparkle button in the inbox composer writes a suggested reply into the box for a human to read, edit and send. Nothing is sent by pressing it.
  • It drafts as the agent that owns the thread, falling back to whichever agent covers the channel.
  • It works even while the agent is switched off — a human is pressing a button and reading the result, which is exactly what a half-built agent is for.
  • Out of the 24-hour window the button is disabled: a reply that cannot be sent is not worth generating. The tooltip says Reply window closed — nothing to draft until this person messages again.
  • Every writing tool is withheld while drafting. Pressing the button three times cannot book three meetings or send three payment requests. Reading tools stay, so a draft can still quote open Cal.com times or a basket link.
  • A draft is written with the plain-text formatting rules, even on a WhatsApp thread. Add any *bold* yourself before sending.
  • If the workspace is out of credits, the error offers a top-up directly.
See Sending messages.

A sensible order to trial one

1

Test drawer

Ask it the ten questions your customers actually ask. Check the sources under each reply — a wrong answer is usually a missing document, not a wrong instruction.
2

Drafts only

Leave Answer inbound messages automatically off and let your team use the draft button for a few days. You see its wording against real messages, with a person between it and the customer.
3

Test mode

Switch automatic replies on with test mode and your own number listed. This is the first time it composes and sends by itself.
4

Live

Turn test mode off. Watch the reply and handoff counts on the agent’s row, which cover the last 30 days.