> ## Documentation Index
> Fetch the complete documentation index at: https://docs.keloa.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Test an agent

> Chat with your agent before letting it loose on customers.

Never let an untested agent go live on a channel. Keloa gives you a private chat panel against any agent — same knowledge, same prompt, same tools — before a real customer ever talks to it.

## Open the Try panel

**AI agents** → click the agent card → **Try this agent**.

You get a chat interface. Ask anything a customer would. Responses show:

* the reply the agent would send
* which knowledge sources it pulled from (click **Sources** to expand)
* any tools it called and the tool result

Nothing here touches a real customer. It's a sandbox.

## What to test

Run at least these five buckets:

| Bucket                    | Examples                                                               |
| ------------------------- | ---------------------------------------------------------------------- |
| **Top 10 real questions** | Past inbox threads; paste them verbatim                                |
| **Edge cases**            | Misspellings, multi-language, long rambles                             |
| **Out-of-scope**          | *"Who is the CEO?"*, *"What's 2+2?"* — should acknowledge and hand off |
| **Angry customer**        | *"This is the third time I've asked…"* — should acknowledge first      |
| **Handoff trigger**       | *"Can I talk to a human?"* — should immediately hand off               |

## What a good answer looks like

* Grounded in a cited source, or honestly says it doesn't know.
* In the customer's language.
* Concise. Three paragraphs is usually too long.
* Offers a next step if it can't fully answer.

## What a bad answer looks like

* Invents a fact. → Add the right answer to knowledge.
* Too formal/stiff or too casual. → Adjust the voice section of the prompt.
* Rambles. → Add *"Reply in one short paragraph unless a list is clearly better"* to the prompt.
* Ignores a clear handoff signal. → Add *"If the user asks for a human, hand off immediately"* to the rules.

## Before you flip to live

Checklist:

* [ ] Tested top 10 real questions
* [ ] Tested "angry customer" and "ask for human"
* [ ] System prompt has a handoff clause
* [ ] Knowledge is current (no stale prices or hours)
* [ ] A teammate is set to receive handoffs

When all five are done, flip **Status → live** and attach to the channel.

<Tip>
  Do the same test after any significant knowledge update — a big crawl can shift answers in unexpected ways.
</Tip>