Testing Agents

The SipPulse AI platform offers the Agent Playground — a unified testing environment where you can interact with your agents via text and voice in the same interface. Validate your agents here before deploying them to production channels like WhatsApp, SIP, Telegram, or web widgets.

Accessing the Agent Playground

1. Navigate to Agents and select an agent
2. Click on the agent name or ⋮ → Open Playground

---

Agent Playground

The Agent Playground combines text and voice interaction in a single interface. You can switch between modes seamlessly during a conversation without losing context.

Interface Layout

Region	Location	Function
Thread List	Left panel	Conversation history with pagination and trigger filters
Message Flow	Center	Where you interact with the Agent via text or voice
Control Bar	Top section	Manages conversation (new, refresh, close, reopen, info)
Information Sidebar	Right panel (toggle with ℹ️)	Context metrics, costs, status, model and tools

Control Bar

Icon	Action	Notes
➕	New conversation	Creates a blank thread
🔄	Refresh	Reloads messages
⏹	Close conversation	Marks as Closed. Runs Post-Analysis if configured
▶️	Reopen conversation	Appears only when closed
ℹ️	Information	Shows/hides the Information Sidebar

Switching Between Text and Voice

The Playground supports seamless mode switching:

• Text mode: Type messages normally in the input field
• Voice mode: Click the microphone button to start a voice session

You can switch between modes at any point — the conversation context is preserved. When you start a voice session, the interface shows a microphone visualizer and displays your speech transcription alongside the agent's responses.

TIP

Tool calls made by the agent are displayed in real time during voice sessions, so you can see exactly what actions the agent is taking while it speaks.

Voice Mode

To use voice mode, your agent must have a TTS model and voice configured in the Voice Settings tab.

Element	Function
Microphone Button	Start/stop voice session
Microphone Visualizer	Visual indicator when your mic is active
Agent State	Shows Idle, Listening, Thinking, or Speaking
Transcription	Your speech and agent responses appear as text in the message flow

Voice session flow:

1. Click the microphone button to connect
2. Speak normally — your speech is transcribed automatically
3. The agent responds with audio and text simultaneously
4. Tool calls appear in real time as the agent processes
5. Click the microphone button again to disconnect

INFO

The agent uses sentence-level TTS streaming, meaning audio playback starts before the full response is generated. This reduces perceived latency.

Thread List and History

The left panel shows your conversation threads with:

• Pagination for navigating through many threads
• Trigger filter to show threads by origin (UI, API, WhatsApp, SIP, etc.)
• Status badges indicating the thread state (Pending, Active, Running, Closed)
• Loading indicator when switching between threads

You can also access your full conversation history from the History page in the sidebar.

Information Sidebar

The Sidebar displays a snapshot of the Agent's configuration at the moment the conversation was created. Later changes are not applied retroactively.

Field	What it shows
Triggered by	Origin channel (UI, API, WhatsApp, SIP). External channels make the input field read-only.
Status	Current thread state
Model	LLM used in this conversation
Context	Tokens used / model limit
Total cost	Input + output token cost
Speed	Average tokens/s during generation
Tools	Available tool calls for this conversation
Post-Analysis	Structured result, if configured and conversation is closed

Additional Instructions & Variables

Before the first message you can add additional instructions that are concatenated to the System Instructions for that conversation only.

Usage examples:

• Pass temporary context: "This conversation is about order #12345."
• Personalize the conversation: "You are talking to Mariana."

If the agent has variables in its instructions, fill in the values before starting the conversation.

INFO

In voice mode, changes to session settings only take effect before connecting. Disconnect first if you need to modify them.

---

Testing Best Practices

General

1. Test before production — cover all critical scenarios, exceptions and alternative flows.
2. Compare models — some tasks work better on larger models; others require lower latency or cost. Create separate conversations for each model and evaluate.
3. Monitor context and cost — check if the Agent exceeds 75% of the model's token limit; simplify instructions when necessary.
4. Validate tools — ensure each tool call returns the expected format and that the Agent handles errors (timeouts, empty responses, etc.).
5. Document learnings — record failures and applied adjustments to build a history of improvements.

Text Testing

• Test error scenarios (empty responses, timeouts).
• Validate behavior with different types of input.
• Use additional instructions to simulate specific contexts.

Voice Testing

• Test in a quiet environment for better speech recognition.
• Validate the agent's response latency.
• Test different accents and speaking speeds.
• Confirm that the selected TTS voice is appropriate for the use case.
• Verify behavior when you interrupt the agent mid-response.
• Check that tool calls display correctly during voice sessions.

---

Post-Analysis

When closing a conversation, the system automatically executes the structured analyses configured in the Agent. The result appears in the Sidebar for consultation.

For more information on structured analyses, see Structured Analysis.

---

Related Resources

• Agent Configuration — Complete reference for profile, instructions, tools and parameters.
• Structured Analysis — How to create automatic metrics for each conversation.
• Agent Deployment — Connect your Agent to WhatsApp, HTTP API and more.
• Conversation Monitoring — Track conversations in real time.