The Playground tab runs the agent against any review you paste, using the same instructions and knowledge it uses live. Nothing here is posted, so there is no way to get it wrong.
What you need: a saved agent with some instructions and at least 1 knowledge source ticked. The playground runs the agent as it is saved, so save your changes first if you want to test them.
Run a review through it
- Open the agent and go to the Playground tab.
- Paste a real review from your feed into Sample review.
- Press Test this review.
- Read the result panel.
The playground showing a sample review, the reply the agent would write, and the knowledge it came from.
The outcomes
- Would send this automatically. The agent is in auto-reply mode, the review cleared every exclusion, and the reply underneath is what would have posted.
- Would wait as a draft. The answer is written and a person reads it first.
- Would leave this review alone. The agent skipped it, and the reason line says exactly why.
- Held by a guardrail. The review asked about the AI itself, asked for the agent's instructions, or tried to give it new ones. The agent stays quiet and leaves it to a person, and the test costs nothing.
The playground reads the text only. It does not take a star rating, so it tests what the agent writes, not which rating band a review falls in. See Rating bands for that.
Reading the reason line
The reason under the result names 1 cause:
- Excluded: the review contains "refund". Your own skip list caught it, and it names the word.
- Your detection prompt reads this as a review that expects no answer.
- Answered from the knowledge pair and its question. The pair is named in a badge above the reply.
- Nothing in the sources this agent reads covered this, so the agent hands it over instead of guessing.
A playground result showing a review the agent would leave alone and the reason why.
That last one is the one to act on. Add a pair to a set in the Knowledge bank, tick that set on the Knowledge tab, and the agent answers it properly next time.
With 2 or more agents, the panel also says which agent the router would pick for this review.
What to test
Test the reviews that worry you, not the easy ones.
- A 1-star complaint. Confirm the agent apologizes and offers a next step, and does not argue.
- A review that names a staff member. Confirm the reply does not discuss them.
- A review in another language, to see the reply come back in the same one.
- A review that mentions a refund, to confirm your skip keyword catches it.
- A question you are certain is covered, to confirm the right pair matched.
Limits
The playground runs at most 20 tests an hour for one agent. When automated replies are paused for the account, a test cannot run the AI either, and the panel says so.
It works on 1 review at a time. It does not tell you how many reviews an agent will answer in a week. For volume, put the agent in draft mode for a week and count what lands.
What this costs
A playground test runs the real AI, so it costs credits like a draft does, and the panel says how many each test used. A test caught by a guardrail or a skip keyword costs nothing. Nothing is ever posted from here.