First Conversation

The first conversation is a controlled test of the configured agent. Run it before sharing a public link, installing the widget, or connecting WhatsApp, Telegram, or another integration.

Open a new session

Open the Hub at /hub. If your organization uses collaborative rooms and the Hub opens a room first, select Agents in the navigation. You can also open Agents in Admin and choose Chat on the agent you want to test.

An empty session shows How can I help?, the Type your message… field, and agent shortcuts under Or talk to an agent. Choose an agent through one of these shortcuts or through the selector in the composer.

The model selector appears only for users allowed to change the model. Keep the agent’s configured model for the first test. If you choose another one, the session uses and saves that selection for subsequent messages.

Prepare a safe message

Start with a short, representative question, such as a common customer request. Connected tools are real: a tool can send data or change an external service. Use disposable data and read-only operations in tests; do not request an irreversible action merely to confirm that a tool exists.

You can attach up to five files. The composer accepts JPG, PNG, GIF, and WebP images; PDF; TXT, Markdown, and CSV; Word; and Excel. The final file must be no larger than 4 MB. Images up to 10 MB and PDFs up to 25 MB may be accepted as inputs for compression. Audio, image generation, and other controls appear only when the agent and organization permissions allow them.

Type the message and press Enter to send it. Use Shift+Enter to insert a line break without sending.

Understand the first send

The screen displays your message immediately and starts streaming the response. The conversation is created only on the first valid send; the URL then changes to the created session. Its title starts as New conversation and may be renamed automatically from the first message.

While the agent responds, its path may show reasoning, tools, and usage. These details help explain the result, but they do not replace checking the response and any external effect.

Passing criteria

Pass the first test only when a response from the selected agent is visible, nonempty, remains in the conversation after a page refresh, and contains no error step. Also check that:

it followed the prompt’s tone and instructions;
it used knowledge-base facts without fabricating information;
it asked for clarification when the request was ambiguous;
it called the correct tools with the expected arguments;
it clearly declined or redirected requests outside its scope;
its length was proportional to the question.

A tool marked as completed does not, by itself, prove that the external effect was correct. When the test involves a safe action, verify its result in the destination system.

If the response fails

Agent unavailable or permission denied: confirm that the agent is visible to your organization and that your role can access the Hub.
Session expired: sign in again and repeat the test in a new conversation.
Insufficient credits: check the organization’s balance before resending.
Connection lost or stalled response: wait for the error message and try a new session. If the stream ends empty or incomplete without a warning, refresh the page and inspect what was persisted. Do not pass the test or repeat an action with external effects before completing that check.

Adjust and retest

If the result does not meet the criteria, return to the editor at /admin/agents/:id/edit and review the prompt, model, knowledge bases, or tools. Then open a new conversation: the previous session retains its own history and can hide the effect of your changes.

Optional next step

After the agent passes in the Hub, you may connect it to a channel. This is optional and is not part of creating or initially validating the agent:

Web chat — public link or site widget.
Official WhatsApp — Cloud API.
Telegram — a bot created through BotFather.
API/Webhook — custom integrations.