Playground

The Playground is a live chat where you talk to your agent exactly the way a customer would, so you can check its answers, watch the tools it calls, and read how it reasons before you put it in front of real people.

What the Playground is for

Think of the Playground as a private rehearsal space. You send messages, the agent responds in real time, and nothing you type here is shown to customers. It is the fastest way to confirm that a change you made to an agent or your knowledge actually does what you expect.

Use it to:

• Try out typical customer questions and see how the agent replies.
• Confirm the agent pulls the right answers from your knowledge.
• Check that tools and integrations fire correctly.
• Catch gaps or wrong answers before they reach a real conversation.

Starting a conversation

Open the Playground from the dashboard sidebar, then:

1. Pick the Agent you want to test from the toolbar at the top.
2. Choose which Version to run (see below).
3. Type a message in the input box and send it, just like a customer would.

Tip: If you have not built an agent yet, the Playground will prompt you to create one first. See Create your first agent to get started.

Testing a specific version

When an agent has more than one version, a Version picker appears next to the agent selector:

This lets you compare a draft against what is live before you publish it. By default the Playground opens on your draft if one exists, since that is usually what you are here to test.

You can also pick a Format preset to preview how the agent's replies look in different output styles.

Resetting the conversation

Each test conversation builds on the messages before it, so the agent remembers earlier turns. When you want a clean slate, click Reset in the top-right corner. This clears the chat, drops the current session, and lets you start fresh without any leftover context.

Note: Switching the agent or version also starts a new conversation automatically, so you are always testing against a clean state.

Inspecting tool calls and reasoning

The Playground does not just show the final answer. It surfaces what happened behind the scenes so you can debug an unexpected reply.

• Tool calls appear as expandable cards in the conversation. Open one to see the exact input the agent sent to a tool and the output it got back, including any errors. This is the quickest way to tell whether a tool or integration misfired.
• Reasoning is shown in a collapsible panel above an answer, so you can read the agent's thinking before it committed to a response.
• Sources behind an answer appear as citation pills in the reply and in the sidebar. Click Review sources under a message, or open the Sources tab, to see which knowledge articles the answer was based on.

The sidebar also has:

• A Session tab with the session ID, the model used, and token counts for the turn.
• A Contact tab showing the simulated test contact and any attributes the agent set during the run.

Picking up where you left off

Past Playground sessions are saved. Click History in the top-right to reopen an earlier conversation, complete with its tool cards and reasoning restored exactly as you saw them live. This is handy for re-checking a tricky case after you have made a fix.

Iterating before you publish

The Playground is meant to be looped through quickly. A typical cycle looks like this:

1. Ask the agent a question in the Playground.
2. Spot a wrong answer, a missing source, or a tool that misfired.
3. Adjust the agent or add to your knowledge.
4. Reset and test the same question again.

When you find a turn worth keeping as a benchmark, click Save as eval under the agent's reply. This captures the question and the expected answer into an eval set so you can re-run it automatically later.

Tip: The Playground is great for quick, hands-on checks. When you want repeatable, measurable testing across many cases, move those saved cases into evals.

Once the agent answers reliably in the Playground, publish it and roll it out to your live channels.