Your agent
Testing in the Playground
Chat with your agent privately, compare different instructions side by side, and review every test session afterwards.
The Playground is where you try your agent before your customers do. It runs the same agent your channels use — the same knowledge, the same actions and the same saved settings — so what you see here is what customers get.
Open it from Playground in the sidebar.
Chat with your agent
Send a message
Type a question in Ask your agent something… and press Enter to send. Use Shift + Enter for a new line.
Read the reply
The agent shows Thinking… while it works, then replies. Answers come from your knowledge sources, and the agent can use your actions just as it would with a customer. Click Stop to cut a reply short.
Start fresh
Click Clear conversation to begin a new session. The old one isn't lost — it stays in your past sessions.
See what the agent did
When the agent uses an action, the reply shows a small chip naming the action and what happened. For example:
| Chip | What it means |
|---|---|
| Collect lead · Saved Alex | The agent captured lead details. |
| Collect lead · Still collecting details | The agent is still asking for required details. |
| Escalate to a human · Handed to a human | The agent would hand this chat to your team. |
| Wrap up conversation · Asked if there's anything else | The agent checked whether the customer needs anything more. |
| Running… | The action is still in progress. |
| This action failed. | The action ran into a problem. |
Booking actions show each step too, such as how many times were offered, Awaiting confirmation and the time that was booked.
Compare instructions side by side
You can run two or more versions of your agent at once, each with its own instructions, and send them the same questions.
Add a variant
Click Compare a variant. A second card appears beside the first, named Variant 1. Give it a clearer name in Variant name if you like.
Edit its instructions
In the variant's card, click Show instructions and change the text in Instructions. The card's subtitle changes to Running edited instructions and a Draft badge appears. Click Reset to saved to go back to your saved instructions.
Ask both the same question
Send the same message to each card and compare the answers.
Keep the winner
Edited instructions are only tried out in the Playground. To use them for real, copy them into Settings → AI and click Save changes. Remove a card you no longer need with Remove variant.
Only the instructions can be changed per variant. Tone, knowledge grounding and every other setting come from Settings → AI. Cards and draft instructions reset when you reload the page.
Past sessions
Every Playground session is saved to your inbox so you can reread it later or share it with a teammate. Click Past sessions to open the inbox filtered to Playground conversations.
Playground sessions carry a Playground badge and are read-only: you can't reply in them, and they never reach a customer. You can also find them from the inbox's channel filter by choosing Playground.
What to know before you test
- Each reply counts toward your plan. The Playground uses the real agent, so every reply uses one of your plan's AI replies. If your allowance is used up, the Playground tells you, and you can upgrade from Settings → Billing.
- The AI rate limit applies. It's counted per teammate in the Playground. If you send too quickly, you'll see "You're sending messages too quickly. Wait a moment and try again." See Auto-reply and rate limits.
- It works while auto-reply is off. Pausing your agent in Settings → Security stops replies to customers, not the Playground, so you can keep testing.
- Some actions are simulated, others run for real. Collect Leads and Create case look successful in the Playground but save nothing. Actions that reach another app — Slack alerts, Calendly and Cal.com bookings, Intercom tickets and custom actions — really run.
Testing a booking action in the Playground can book a real meeting, and a Slack action posts a real alert. Point those actions at a test calendar or channel while you experiment.
Troubleshooting
The agent says it doesn't have that information
It couldn't find a relevant passage in your knowledge. Check that the right source exists and shows Trained, or add a Free text answer. If your grounding is set to Strict, the agent won't fill gaps from general knowledge.
"The agent couldn't reply. Please try again."
Something went wrong with that turn. Send the message again. If it keeps happening, check that your plan is active and has AI replies left in Settings → Usage.