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

# Test conversations

> Check an agent's draft against made-up callers before you publish it.

A test call from the **Test** button shows one conversation: the one you
happen to have. The **Tests** tab in the agent editor checks several at once.
The [draft](https://docs.brosbots.com/en/voice-agent/draft-publish.md) agent talks in writing to made-up
callers, and an AI model judges each conversation.

## Scenario

A scenario is one made-up caller:

| Field | What to enter |
| - | - |
| **Name** | Short, e.g. "Visit tomorrow morning" |
| **Caller** | Who they are, how they talk and which details they give when asked, e.g. "Anna Smith, calling in a hurry, wants a visit tomorrow before 10:00, gives her number only when asked" |
| **Goal** | What should be true at the end of the call, e.g. "A visit booked for tomorrow before 10:00 and confirmed" |
| **Details to collect** | Fields the agent should take from this caller (key and description) |

An agent can have at most 10 scenarios. On an empty tab, **Suggest 3
scenarios** writes three from the agent's instructions: a typical caller, a
hurried caller who gives details out of order, and a caller with an unusual
request. It takes the fields to collect from the ones the agent already
collects. You can edit or delete the scenarios afterwards.

## Running and the result

**Run all** plays every scenario; the ▶ icon next to a scenario plays only
that one. A run usually takes from a few seconds to a minute. The caller has
at most 12 turns.

A scenario **passed** when the goal was reached and the agent collected every
field on the list. Expand it to see the verdict, each field with its value or
"not collected", and the whole transcript.

A result is about the draft that was saved when the run started. After any
change in the editor, an old result is marked **Result from an earlier
version**: run the tests again to check the current draft. Without a draft,
the published agent is tested.

**Could not be played** means an error, not a verdict, e.g. no API key for
the model provider the agent uses.

## Publishing

When a scenario on the current draft failed or could not be played,
**Publish** asks first whether to publish anyway. Tests never block
publishing.

## What tests do not check

* **The voice.** The conversation is written: no pronunciation, latency or
  interruptions.
* **Tools.** The agent does not really book, send an SMS or transfer the
  call. In the transcript it writes in square brackets what it would do,
  e.g. "\[booking a visit: tomorrow 10:00, Anna Smith]", and assumes it
  worked.
* **Transitions between flow stages.** The agent gets the global prompt, the
  greeting and every conversation stage in the order a call walks them.

The verdict is an AI model's reading of the transcript, so treat it as a
hint. When a result surprises you, read the transcript.

## Limit

An agent's tests can be run 20 times a day. Running all scenarios counts as
one. Test conversations use no package minutes and no wallet balance. When
the agent uses its own model provider key, its replies go on that key.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.