Playground & Chat

The Playground puts you in the customer's seat. Type what a customer would say, on the channel they'd say it, and see the exact reply your agent would produce - plus whether it would send that reply by itself, hold it for approval, or hand the conversation to a human.

Testing your agent#

1Open Agent → Playground and pick the agent you want to test.
2Choose a channel: Website, Email, WhatsApp, Instagram or Voice.
3Type a message as the customer - or click one of the built-in scenarios, which cover a normal question, a pricing question, a booking request, a discount push, and a request for other customers' data.
4Read the reply, the verdict chip, and the signals underneath it.

This is the production reply engine, not a preview of it. Your agent's instructions, tone, rules, knowledge base, custom tools, confidence threshold and escalation triggers all apply exactly as they would to a real customer.

Playground spends credits

Every message is a real model call, so it debits credits like a live conversation. With an empty balance the Playground returns insufficient_credits instead of a reply.

Reading the verdict#

The chip above each reply is the answer to "what would actually have happened?".

Chip	Meaning
Would auto-send	Confidence cleared your threshold and that channel's auto-reply is on. In production this reply goes out with nobody looking at it.
Held for review	Confidence was below your threshold. The reply waits in the inbox as a draft.
Not connected	Confidence was high enough, but auto-reply is off for that channel - so it would still be a draft. Fix it in the agent's Autopilot sidebar, not here.
Escalated to human	An escalation trigger fired - the agent stops replying and hands the conversation over, regardless of confidence.

The explanation next to the chip always names your actual threshold, so a "held" verdict tells you how far short it fell rather than leaving you guessing.

Signals#

Under each reply is a short list of what shaped it:

Which tone was applied, and how the channel changed the shape of the message.
"Answered from your knowledge base" - shown only when retrieval actually found something. Its absence on a factual question means the agent answered from its business description, which is your cue to add that fact to the knowledge base.
The escalation reason, when one fired, plus a note that a human would be brought in.
Which follow-up day would be scheduled if the customer went quiet.
A warning when the channel's auto-reply is off.

How channels differ#

Channel	What you get back
Website	Quick conversational chat, the same engine the widget uses.
Email	A full email: subject line, body and sign-off.
WhatsApp / Instagram	Short, casual messages shaped for a chat thread.
Voice	Spoken-style phrasing, rendered through the real voice prompt builder so it reads like what a caller would hear rather than an email read aloud.

Each channel keeps its own thread

Switching channels doesn't clear anything - Website, Email, WhatsApp, Instagram and Voice each hold their own conversation, so you can compare the same question across all five without retyping it.

Sessions & history#

Every Playground session is saved and reloadable from the history panel, grouped by date and marked per channel. Reopening one restores the messages along with their verdicts and signals. Clearing a channel starts a fresh session without touching the others, and any session can be deleted outright.

Copilot chat#

Copilot (Dashboard → Chat) is the other half: an assistant for you, not a simulation of your customer-facing agent. It can read and act on your workspace through tools rather than just talking about it.

Ask it to	What it uses
Show or update tasks, add comments	Tasks
List, book, reschedule or cancel meetings	Calendar
Look through conversations and drafts, approve or reject a draft	Inbox
Find and update leads, move pipeline stages	Leads
Summarise performance	Analytics
Search across the workspace, check who's on the team	Search, Team
Set up or adjust an agent, query the knowledge base	Agent

Copilot obeys your permissions - it can only touch what your role lets you touch, and only the agents you have access to.

What it doesn't do#

Nothing is sent. No email leaves, no WhatsApp message is delivered, no call is placed.
No lead, task or conversation is created in your real inbox from a Playground message.
Voice is typed, not spoken - it shows you what the agent would say, not how the speech pipeline handles interruptions. Use a test call from Agent → Voice for that.
Follow-ups are described, not scheduled.

A good test loop

1Run the five built-in scenarios on your busiest channel.
2Anywhere the knowledge signal is missing but should be there, add that fact to the knowledge base and re-ask.
3Push on the guardrails - ask for a discount, ask for another customer's details - and confirm the agent refuses or escalates.
4Only then turn on auto-reply for that channel.

PreviousSetting Up Your Agent NextKnowledge Base