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

# Testing agents

> Use the in-browser simulator to validate behavior before publishing.

Every iteration on an agent should pass through the **Tester** before going live. The tester runs the agent's draft version end-to-end with a browser-based microphone, so you can hear exactly what callers will hear.

## Open the tester

On the agent edit page, click **Tester** (top-right). A side panel opens with a microphone control, a live transcript, and a tool-call log.

## Test loop

<Steps>
  <Step title="Click Start call">
    The browser will request microphone access. The agent's draft version is loaded.
  </Step>

  <Step title="Talk">
    Speak as a caller would. Try the happy path first, then push on edge cases.
  </Step>

  <Step title="Watch the transcript">
    The transcript updates in real time. Tool calls show inline with their inputs and outputs.
  </Step>

  <Step title="End the call">
    Click **End call** when done. The transcript persists in your test history for the session.
  </Step>
</Steps>

## What to test

A good pre-publish test plan covers:

| Scenario           | What you're validating                                       |
| ------------------ | ------------------------------------------------------------ |
| Happy path         | Agent achieves its primary goal                              |
| Off-topic question | Hard rules hold; agent doesn't hallucinate                   |
| Interruption       | Agent recovers gracefully when you talk over it              |
| Silence            | Agent's silence-handling behavior fires correctly            |
| Tool failure       | If the workflow tool errors, the fallback message plays      |
| KB miss            | When you ask something not in the KB, agent admits ignorance |
| Spelling/numbers   | Agent collects names, emails, phone numbers without garbling |
| Hangup             | Agent ends the call cleanly when the conversation is done    |

## Reading the tool-call log

Each tool invocation logs:

* **Tool name** and the **arguments** the LLM filled
* The **action sequence** that ran
* Any **error** with the failing step

If the agent calls the wrong tool, the fix is usually in the tool description (the LLM read it and got the wrong cue). If the agent calls the right tool with bad arguments, the fix is in the parameter descriptions.

## Tester vs real call

The tester is faithful to almost everything but:

* Phone audio quality (compressed, \~8kHz)
* Real-world background noise
* Cellular jitter

For high-stakes agents (sales, medical), always also place a few real calls from your phone before publishing widely.

## Iterating fast

A good loop:

1. Edit the prompt or a tool
2. Click **Test** without leaving the page (the tester picks up draft changes immediately)
3. Run the same 3-5 scenarios you ran last time
4. Save and repeat

Don't republish between every iteration — the published version is what real callers hit.