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

# Testing Best Practices

> A master guide to ensuring your Voice AI Agents are production-ready and error-free.

# End-to-End Testing Strategy

Deploying a **Voice AI Agent** is different from deploying a Chat Agent. You aren't just limited to testing logic but testing acoustics, latency, and patience. A script that reads well on screen might sound robotic, too long or rushed over the phone.

This guide outlines the **Golden Path** for testing—from the first draft to post-production updates—ensuring your **Native English Voice Agent**, **Voice Agents for Indian Languages**, **Arabic Accent AI Agents** and Voice Agents in 80+ other languages perform flawlessly.

***

## Phase 1: The Build Phase (New Agents)

When building an agent from scratch, your goal is to validate **Logic** before **Voice**.

<Steps>
  <Step title="Logic Validation (Chat Simulator)">
    **Tool:** [Chat Simulator](/test-voice-agent/manual-chat)

    Before worrying about accents, ensure the **AI Agent** follows your rules.

    * **Happy Path:** Test the ideal customer journey (e.g., User says "Yes" -> Agent books appointment).
    * **Unhappy Path:** Test rejection (e.g., User says "No, I'm busy" -> Agent handles objection or hangs up).
    * **Context Check:** Use the **Debug Mode** to ensure variables like `user.name` or `lead_status` are being captured correctly from the start.
  </Step>

  <Step title="Acoustic Check (Web Call)">
    **Tool:** [Web Call Testing](/test-voice-agent/web-call)

    Once the logic holds, test how it sounds without spending money on telephony credits.

    * **Speed & Tone:** Does the agent speak too fast? Is the selected voice (e.g., *Riya* or *Aditya*) too formal for a sales call?
    * **Dialect Verification:**
      * **For Arabic:** Speak in a specific dialect (e.g., *Khaleeji*). Does the agent understand? If not, add **Boosted Keywords**.
      * **For Hinglish:** Speak a mixed sentence like *"Mera loan application approve hua kya?"*. Verify the transcription in the live chat log.
  </Step>

  <Step title="Pronunciation Tuning">
    **Tool:** [Word Management](/build/configure-voices/add-pronunciation)

    If the agent mispronounces your brand name during the Web Call:

    1. Go to **Word Management**.
    2. Add the phonetic spelling.
    3. **Retest immediately** via Web Call to verify the fix.
       <Warning>Only add words that are critical. Overloading this list adds latency.</Warning>
  </Step>
</Steps>

***

## Phase 2: The Staging Phase (Pre-Production)

Before you assign this Recipe to your main business number, you must test against **Real World Friction**.

### 1. The "Background Noise" Test

**Method:** Call the agent using the [Real Phone Call](/test-voice-agent/phone-call) method while standing in a noisy environment (or play cafe noise in the background).

* **Goal:** Test **Interruption Sensitivity**.
* **Fix:** If the agent stops talking every time a car honks, lower the sensitivity in your Global Settings.

### 2. The "Latency" Test

**Method:** Call from a mobile network (4G/5G), not Wi-Fi.

* **Goal:** Feel the delay.
* **Benchmark:** \* **Good:** \< 1 seconds response time.
  * **Bad:** > 3 seconds.
* **Fix:** If latency is high, shorten your **System Prompt** or remove unnecessary complex logic blocks at the start of the flow.

### 3. The "Silence" Test

**Method:** Stay silent when the agent asks a question.

* **Goal:** Verify the **End Call on Silence** triggers correctly (e.g., after 1 minute) or that the agent prompts you again ("Are you still there?").

***

## Phase 3: Making Edits (Regression Testing)

When you need to update an existing live agent (e.g., changing the pricing or adding a new holiday greeting), follow this strict protocol to avoid breaking production.

<CardGroup cols={2}>
  <Card title="1. Clone & Edit" icon="copy">
    Never edit the live Recipe directly. **Clone** the Recipe, make your changes, and test on a temporary number or Web Call.
  </Card>

  <Card title="2. Regression Test" icon="rotate-left">
    Test the *unchanged* parts of the flow. Did adding a new "Holiday" block accidentally break the "Transfer to Agent" logic?
  </Card>

  <Card title="3. Context Simulation" icon="flask">
    Use **Web Call Custom Configuration** to inject variables. If your edit changes how "Premium Users" are handled, manually inject `account_type: premium` in the test setup to verify the new path.
  </Card>

  <Card title="4. Hot-Swap" icon="shuffle">
    Once validated, simply switch the **Phone Number** configuration to point to the new Recipe version. This ensures zero downtime.
  </Card>
</CardGroup>

***

## Production Monitoring Checklist

Even after deployment, your job isn't done. Use **Post-Call Analysis** to automate quality assurance.

<Check>
  **Daily:** Check **Post-Call Analysis** logs for "Unknown" intents. If users are asking for something you didn't account for, add a new path.
</Check>

<Check>
  **Weekly:** Review **Speech Recognition** logs for low-confidence transcriptions. Add these terms to **Boosted Keywords**.
</Check>

<Check>
  **Monthly:** Re-evaluate your **Voice Profile**. New, higher-quality voices (e.g., ElevenLabs Turbo) are frequently added to the library. Swapping to a newer model can instantly decrease latency.
</Check>

***

## Summary: The Testing Pyramid

| Layer      | Method              | Focus                        | Frequency            |
| :--------- | :------------------ | :--------------------------- | :------------------- |
| **Top**    | **Real Phone Call** | Latency, Network, Noise      | Final QA only        |
| **Middle** | **Web Call**        | Accents, ASR, Pronunciation  | Weekly / Major Edits |
| **Base**   | **Chat Simulator**  | Logic, Prompts, Data Capture | Daily / Every Edit   |

By adhering to this pyramid, you ensure that 90% of bugs are caught in the cheap, fast **Chat** layer, leaving the **Phone** layer for final polish of your **Voice AI Agents**.