Test an agent in the Playground

What the Playground is

The Playground is a test chat for your agent. Open Playground in the sidebar of any agent to try questions the way a customer would. Answers come from the same knowledge, instructions, and behavior as your live agent, so what you see is what a customer would get. Questions about dates and times, like "are you open tomorrow?", resolve in your own timezone, just as they do for a customer.

Every visit starts a fresh test session. Test conversations are saved in your Inbox on the playground channel so you can review them later, and they are never shown to customers.

Run a test

1. Open your agent and select Playground in the sidebar.
2. Ask a question, or pick one of the suggested starters. Answers that include images, audio, video, or files show that media in the Playground, just like they do in the chat widget.
3. Select an answer to open the inspection panel. It shows whether the answer was AI-generated or an Instant Answer, how long it took, and every source it was based on.
4. Judge the answer:
   • Looks good marks it as reviewed, so you can keep track during a longer test run.
   • Fix answer opens a form pre-filled with the question and the agent's reply. Write the answer you want, save it, and it becomes an Instant Answer that future matching questions will use.
5. After saving a fix, use Retest question to ask the same question again and confirm the corrected answer comes back.
6. Use New session to start over with a clean conversation. Earlier context can influence answers, so retest important questions in a fresh session too.

Build a small test set

Include questions with clear source answers, alternate wording, multi-part questions, unsupported requests, and facts that recently changed.

If you use commerce, include a product search and an order lookup. If you use Functions or Guided Flows, test their success and failure paths.

Success criteria

• Answers match current approved content.
• The agent does not invent an answer when evidence is missing.
• Cited sources open the expected page.
• Lead capture and handoff appear at the expected time.

Repeat this check after major source, instruction, model, integration, or Guided Flow changes.

Test live delivery too

The Playground does not prove that your site or channel is connected. Send one message through every enabled surface and confirm it appears under Inbox with the correct channel.

Troubleshooting

"Could not start a test session." Check your internet connection and select Try again. If it keeps failing, refresh the page or sign out and back in.

The corrected answer does not come back when I retest. Questions worded exactly like your correction apply right away. Reworded versions of the question need a short background update after saving; wait a minute and retest. If it still does not apply, check that the Instant Answer is enabled under Sources.

The agent stopped answering in my test. If your test triggered a handover to a human teammate, the agent stays quiet in that conversation, just like it would for a customer. Start a New session to keep testing AI answers.

Do test messages count toward my plan? Yes. Playground messages use your message allowance the same way customer messages do.

I see my test chats in the Inbox. That is expected. Test conversations are labeled with the playground channel so you can tell them apart from real customers, and you can delete them like any other conversation.

I can't open the Playground. You need permission to view the agent. Ask the agent owner to check your team role.

Repeated knowledge questions

Fetchply may reuse a recent, source-backed answer when a visitor asks the same stable knowledge question again. It can also recognize conservative wording changes such as “How do refunds work?” and “How does a refund work?”. Common Roman Urdu spelling variations are also recognized for stable questions, for example mujhy, mujhe, kaam, kam, karta, and krta.

Personal, live, account, order, booking, pricing, availability, action, and integration requests are always handled normally instead of reusing a shared answer. Follow-up questions that depend on earlier messages are also handled normally.

A reused answer counts as one message toward the agent owner's message allowance, just like any other delivered reply.

Ordinary wording quoted from your setup choices, such as "I don't have a website," does not prevent an otherwise complete answer from being reused. Incomplete or error replies are never reused.

When you retrain the agent or change its knowledge or instructions, previous answers stop being reused automatically. If two questions have different numbers, names, or negative wording such as “does” and “does not,” Fetchply treats them as different questions.