Testing Agents
The test interface is a live debug window that lets you converse with your agents exactly as a real visitor would — with full visibility into every internal decision the system makes.
Go to Agents → Tests.
What the test window shows
The conversation window renders two types of messages interleaved:
- Regular messages — your input and the agent’s responses, exactly as a visitor would see them
- Debug blocks — green-tinted panels between messages, showing internal system events invisible to real visitors
The debug blocks are the core of the tool. They appear automatically after each processing step.
Debug events — step by step
Here is the full sequence of what happens after you send a message, and what debug information appears at each step.
1. Language detection
The system detects the language of your message. The agent will respond in the same language, regardless of what language your documents are written in.
2. Orchestrator routing
The orchestrator analyzes your message and matches it against all agent descriptions and routing examples.
The debug block shows:
Comment: Routing this request to supportbot. This bot starts extracting the missing request details.
If your message is ambiguous between two agents, the orchestrator asks a clarifying question instead of routing — and the debug block explains why.
3. Data extraction
The assigned agent checks which extraction fields are defined and which values are already present in the conversation. If required information is missing, the agent asks the user for it.
The debug block shows the current state of the extracted JSON:
{
"email": "[email protected]",
"incident_details": null
}Fields still set to null are ones the agent will ask for before proceeding.
4. Search query generation
Once required data is collected, the agent formulates a search query. This is not necessarily the exact words the visitor typed — the LLM synthesizes a more precise query based on the full conversation context.
The debug block shows:
Comment: Search query formed: password reset not working after email change
If the agent cannot formulate a clear search query from the available context, it will ask the visitor a follow-up question to clarify intent before searching.
5. Document chunks retrieved
The agent runs the search query against its assigned document categories (both keyword and context search simultaneously) and retrieves the most relevant text fragments.
The debug block shows the actual chunks found:
[
"To reset your password, go to Settings → Security → Reset Password...",
"If you changed your email address, you must verify the new address before..."
]These are the exact pieces of content the agent will use to compose its answer. If the chunks are irrelevant or empty, the agent will either ask for clarification or trigger the fallback behaviour configured for that agent.
6. Response generation
The agent composes its response based on the retrieved chunks and sends it as a regular message.
7. Token counter
The token counter in the top-right of the chat window updates after each full exchange, showing the cumulative token spend for the current test session.
Reset button
Click Reset to clear the conversation and start a new session. This resets the extracted JSON state, token counter, and conversation history. Use it between test scenarios to avoid context bleeding from the previous conversation.
Common things to check
Wrong agent is handling the question → The routing debug block will show which agent was selected and why. Improve that agent’s description or add a routing example that covers the misrouted query.
Agent asks too many questions before answering → Review the extraction fields — remove any that aren’t essential for your workflow.
Retrieved chunks are irrelevant → The document content may not match the search query well. Check document quality in that category, or adjust the agent’s scope.
No chunks found, agent triggers fallback → The knowledge base doesn’t contain an answer for that query. Add a document covering the topic.
WebSocket error on load → Check your token balance and ensure all required fields (short name, description, at least one category) are set for every agent.