> For the full page index, fetch https://echonova.ai/llms.txt # Test your agent Open the playground, talk to your agent by chat or by call, and understand what the test does and does not prove. The playground is a real conversation with your agent, started from the portal instead of from a channel. It runs the same agent, the same model, the same knowledge and the same tools a customer would reach, and it is charged the same way. Use it to hear the agent before you connect anything to it. ## What you need - An agent that is not archived. The **Test** button is disabled while the agent is archived. See [Agent status](/docs/agents/agent-status). - Credits on the workspace. The session is refused before it starts if the balance cannot cover a turn. - A microphone and permission to use it, for **Call** mode only. **Chat** mode needs nothing but the browser. ## Steps ### Step 1: Open the playground Press **Test** in the agent header. It is on every tab, so you do not need to be anywhere in particular. The agent's Playground tab carries the same thing under **Test your agent**, described as **Start a live voice session using the agent's current configuration.** An overlay opens with the caption **Start a test session** and two choices, **Chat** and **Call**. ### Step 2: Choose Chat or Call **Chat** starts a text session. Your browser exchanges text with the agent, with no speech in the loop. **Call** starts a voice session. Your browser sends your microphone and plays the agent's speech back, which is the full pipeline: speech to text, then the model, then text to speech. The caption changes to **Try Chat** or **Try Call** and the status line reads **Connecting…** while the session is being set up. ### Step 3: Talk to the agent In **Chat**, type into the box labelled **Type Your Message Here** and press the send control, **Send message**. While the agent is answering, the send control is disabled: the playground allows one turn at a time. Three dots appear while the agent has your message and has not started replying, labelled **Agent is thinking…**, and the reply fills in as it arrives rather than landing whole. In **Call**, just talk. The status under the controls reads **Listening…** while the agent is waiting on you and **Agent is speaking…** while it talks back. **Mute microphone** stops your audio and the status changes to **Microphone off**. Pressing it again reads **Unmute microphone**. ### Step 4: End the session **Cancel** ends the session and returns you to the **Chat** or **Call** choice, so you can try the other one. The close control, the Escape key and clicking outside the overlay all end the session and close the playground. Closing always ends the session. There is no way to leave one running behind the overlay. ## Get it right A test proves one change at a time. Decide what you are checking before you press **Test**, and keep the same questions from one run to the next. **Recommended** - Write down three questions a customer asks, run them in **Chat** after every change to the instructions or the knowledge, and compare the answers. - Switch to **Call** to check the voice, the pauses and how the agent handles being interrupted, once the answers are right in **Chat**. **Less effective** - Chatting freely, changing the persona, the model and the knowledge together, then testing again. - Running every test as a **Call**. A call costs more than a chat of the same length and tells you nothing extra about the wording. ## Worked example You have just added the returns policy to the agent's knowledge and turned on **Enable retrieval**. 1. Press **Test**, choose **Chat**, and type `Can I return shoes I bought 20 days ago?` 2. If the policy is **Ready** and selected, the answer quotes it. If the agent answers from general knowledge instead, work through [the checklist](/docs/agents/knowledge-troubleshooting). 3. Press **Cancel**, choose **Call**, and ask the same question aloud. Listen for the same facts. 4. Open the **Conversations** tab. Both sessions are there with their transcript and **Total credits**. ## What happens next The session is recorded like any other. A **Chat** test is kept as a web chat conversation and a **Call** test as a web voice one, both on the agent's Conversations tab, with their transcript, their credits and their tool calls. The agent's session count goes up, and a **Call** test counts toward the voice figures on the Overview tab. Nothing about the test is marked as a test. Once it is recorded it is a conversation like the rest. ## The first line you see In **Chat**, the opening line is on screen before the agent has said anything. It is the agent's **Greeting message**, exactly as it is saved on the Persona tab. The agent also greets when it joins. When your greeting is already on screen, that first incoming line is not shown a second time. When the agent has no greeting saved, there is nothing to show up front and the agent's own opening line appears instead, so you see whatever it chose to open with. In **Call**, the greeting is spoken, not written. Nothing is shown before the agent starts talking. ## What the playground costs The same as a real conversation, at the same rates. A call is charged for the time it is connected and every turn for its AI usage, and there is no discount, no exemption and no free allowance for a session started from the portal. See [What a turn costs](/docs/agents/what-a-turn-costs). Two consequences worth planning for: - A **Call** test costs more than a **Chat** test of the same length, because a call also pays 2,500 credits for every minute it is connected. - Leaving a call connected while you read something else is charged for every minute it stays connected. Press **Cancel** when you are done. If the balance runs out mid-call, the session ends, the playground shows **The call ended because the workspace ran out of credits.** and offers **upgrade your plan** beside it. ## What the playground does not exercise The playground runs the agent. It does not run the channel, so everything a channel adds is missing from the test. | Not exercised | On a channel | | ----------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | | The handoff instruction | Helpdesk and Genesys conversations add a handoff instruction to the agent's instructions. The playground uses your instructions as written | | Conversation history | A text channel carries earlier turns into a resumed conversation. A playground chat starts with none | | Contact identity | A channel conversation knows who the customer is from the provider's profile. A playground session has no contact | | Delivery back to the customer | A channel reply is handed to the provider to send, and can fail there. In the playground the reply is already in front of you | | Session close rules | Channel conversations close on a timeout, a human takeover, an escalation or a reassignment. A playground session ends when you end it | | Working hours | A playground session starts at any time of day, as a channel conversation does. See [Working hours](/docs/agents/working-hours) | A clean playground run tells you the persona, the model, the voice, the knowledge and the tools are right. It does not tell you the channel is connected correctly. ## Limits | Limit | Value | Applies to | | ---------------------------------------------------- | ----------------------------------------------- | ------------------------------------------------------------------ | | Time allowed to start a session | 20 seconds | **Chat** and **Call** | | Time allowed for a reply before the turn is given up | 60 seconds | **Chat** | | Turns in flight at once | 1 | **Chat** | | Time allowed to end a session cleanly | 5 seconds | **Chat** and **Call** | | Transcript kept when you reopen the playground | None. Reopening clears the thread and any error | **Chat** | | Testing an archived agent | Refused | **Test** is disabled, and a session cannot be started on the agent | ## Troubleshooting ### No reply. Try again. Sixty seconds passed between your message and the first word of the answer. The turn is given up and this line takes the place of the agent's bubble. The session stays open, so you can send again. A slow tool call and a slow model both look like this. Check the conversation afterwards for a tool call that did not return. ### Could not start the voice session. Or, in chat mode, `Could not start the chat session.` The session could not start. Check, in order: the workspace has credits, and the agent is still active, since it may have been archived in another tab. Then close the playground and try again. If it keeps failing, contact support with the agent id. ### Timeout has occurred The start did not finish within 20 seconds. Close the playground and try again. Repeated timeouts are not caused by the agent's configuration; if they continue, contact support. ### The call ended because the workspace ran out of credits. A turn could not be charged, so the session ended. The message carries an **upgrade your plan** link. Nothing is lost: the turns that were charged are on the conversation. ### The agent speaks but I hear nothing The browser blocked audio playback, or the output device is wrong. Check the browser tab is not muted and the right output device is selected, then reload and start again. ### The greeting appears twice It should not. The playground shows your saved greeting once and does not repeat the agent's own opening line. If you see two, the second one is not the greeting: it is the agent's first real answer. ## FAQ ### Do playground conversations cost credits? Yes, at exactly the same rates as a customer conversation. There is no test mode. ### Do playground conversations show up in my analytics? Yes. They are ordinary conversations, counted in the voice and text figures and in the credits used. ### Can I test an archived agent? No. **Test** is disabled while the agent is archived, and a session cannot be started on it; the refusal reads `Cannot start a session on an archived agent`. ### Does the playground use my saved settings or the ones I just typed? The saved ones. The session takes the agent as it is saved, so save your edits before you test. Changes made during a session apply to the next one. ### Can I see the transcript of a call I just made? Not in the playground. Voice mode shows no words on screen. The transcript is on the agent's Conversations tab once the call ends. ### Why does the chat let me send only one message at a time? Because the agent answers one turn at a time. The send control unlocks when the reply is complete or the 60-second wait expires. ## Related - [Agent status](/docs/agents/agent-status): why **Test** is disabled on an archived agent - [What a turn costs](/docs/agents/what-a-turn-costs): what each test session is charged - [Working hours](/docs/agents/working-hours): what the window does and does not change - [The agent overview tab](/docs/agents/agent-overview-page): where a test session shows up in the numbers - [Write the persona](/docs/agents/persona): the greeting and instructions the test runs - [Choose a model and reasoning effort](/docs/agents/model-and-reasoning): the model the test uses - [Voices and language](/docs/agents/voice-and-language): the voice you hear in **Call**