This article explains how to use Hugo Observability to understand, monitor, and debug individual AI Agent conversations.

Hugo AI Agent can answer customers, use support resources, call tools, trigger workflows, and escalate conversations when needed.

Hugo Observability gives you a detailed view of what happened during each conversation. You can review the steps processed before Hugo answered, inspect the tools and models used, identify errors, and understand why a conversation was handled, escalated, or failed.

In this guide, you will learn how to:

• Understand Hugo Observability → see what happened during individual AI Agent conversations
• Know when to use Observability → investigate unexpected answers, errors, and escalations
• Open the Observability view → access it from your AI Agent setup
• Find a conversation to inspect → filter conversations by outcome
• Read a conversation trace → follow the steps processed by Hugo
• Understand conversation metrics → review tools, steps, and reasoning time
• Review models, tools, inputs, and outputs → understand which resources Hugo used
• Investigate failed or escalated conversations → identify where the conversation needs improvement
• Open the original conversation → return to the Crisp Inbox with the full context
• Export Observability traces → share or investigate conversation traces
• Understand Observability vs Analytics → choose the right view for your analysis
• Follow Observability best practices → turn traces into useful improvements

What is Hugo Observability

Hugo Observability is a conversation-level monitoring view for your AI Agent.

It shows how a specific customer conversation was processed by Hugo or by a Workflow before the final outcome.

Instead of only seeing the customer message and Hugo's answer, you can inspect the steps that happened between them.

Depending on the conversation, Observability can show:

• The tools Hugo used
• The model that generated an answer
• Searches performed before answering
• Tool inputs and returned results
• Authentication steps
• Guardrails applied to the response
• Processing and reasoning times
• Errors generated during the conversation
• Whether the conversation was handled, escalated, processed by a Workflow, or failed

This helps you understand not only what Hugo answered, but also how the conversation reached that result.

When should you use Observability

Use Observability when you need to understand the behavior of a specific AI Agent conversation.

Investigate an unexpected answer

If Hugo gives an incomplete, inaccurate, or surprising answer, review the trace to see which sources, tools, and model were used before the response was generated.

Check whether a tool was called

When Hugo should retrieve customer information or perform an action, Observability shows whether the expected tool was called and what it returned.

This can be useful when investigating Hugo Integrations, MCP tools, or Widget Tools.

Understand an escalation

Review the conversation steps to understand what Hugo tried before handing the customer over to your team.

You can then check whether your Instructions, Routing rules, or escalation setup should be adjusted.

Diagnose an error

If a conversation fails, Observability displays the error inside the trace and helps you identify which step or tool caused it.

Investigate a slow conversation

Review processing times to see whether the delay came from Hugo's reasoning, a search, an external tool, or another step in the conversation.

Share a technical example

Export the trace when you need to share a reproducible conversation with your technical team or Crisp support.

Open the Observability view

To access Hugo Observability:

1. Open Crisp
2. Select AI Agent from the main menu
3. Go to Evaluate → Observability

The view is divided into two main areas:

• Conversation list → find and filter conversations processed by Hugo or Workflows
• Conversation trace → inspect the detailed steps for the selected conversation

Select a conversation from the list to open its trace.

Find the conversation you want to inspect

The Observability list contains recent conversations processed through your AI Agent setup.

You can use the status filter to focus on a specific outcome.

Available filters include:

• All → show all available conversations
• Handled → conversations completed by Hugo
• Workflow → conversations processed through a Workflow
• Escalated → conversations transferred to a human
• Failed → conversations where an error prevented the expected result

Each conversation preview can also display information such as:

• The latest customer message
• The conversation outcome
• The number of tools used
• The total processing time
• When the conversation happened

Start with Failed or Escalated when investigating a recurring problem.

Read a conversation trace

A conversation trace displays the customer messages, Hugo's answers, and the steps processed between them.

The trace follows the conversation chronologically, so you can review what happened before each answer.

Depending on the conversation, a trace can include:

• Agent Reasoning → the reasoning stage processed before Hugo continues
• Tool call → an action or data request sent to an integration
• Tool result → the information returned by that tool
• Search → a search performed across available resources
• Guardrails → checks applied before a response is generated
• Authentication → verification required before accessing protected data or actions
• Workflow steps → actions processed by a Crisp Workflow
• Errors → problems encountered while processing a step
• Hugo answers → the customer-facing messages generated by the AI Agent

Click a step to expand it and inspect the available details.

Understand the conversation metrics

At the top of a selected conversation, Observability displays a quick overview of its processing activity.

The main metrics include:

• Tools used → the number of tool calls made during the conversation
• Steps processed → the number of processing steps completed by Hugo
• Average reasoning time → the average time spent on reasoning stages

These metrics help you understand the complexity of a conversation.

For example:

• A conversation using several tools may require data from multiple systems
• A high number of steps may indicate a more complex customer request
• A slow tool call may explain why the customer waited longer for an answer
• A long trace may reveal unnecessary or repeated actions

These metrics should be interpreted in context. A higher number of steps does not automatically mean that something went wrong.

Review models, tools, inputs, and outputs

Observability lets you inspect the resources Hugo used to answer or perform an action.

Review the AI model

The model name is displayed under answers generated by Hugo.

This helps you confirm which model generated a specific response.

To learn how models affect quality, speed, hosting, and AI credit usage, read How to choose and configure Hugo's AI Model.

Review a tool call

Open a Tool call step to inspect the action performed by Hugo.

The available details can include:

• Tool name
• Tool provider
• Execution time
• Input parameters
• Returned output
• Request and response logs
• Success or error status

This helps you verify whether the expected information was sent to the tool and whether the result returned to Hugo was correct.

Review authentication steps

When an action requires customer verification, the trace can display the authentication step and its outcome.

This is useful when Hugo needs to access private account, billing, or order information before completing a request.

Investigate failed or escalated conversations

Observability helps you move from a conversation outcome to the step that caused it.

Investigate failed conversations

Open a conversation marked as Failed, then review the trace from top to bottom.

Check:

• Which step contains the error
• The error message displayed
• Which tool or provider was involved
• The parameters sent to the tool
• Whether the input matches the expected format
• Whether a valid result was returned
• Whether the external service was available

Once you identify the failing step, you can review the related integration, tool configuration, or Workflow.

For a broader diagnostic checklist, read Troubleshooting and common questions about Hugo AI Agent.

Investigate escalated conversations

Open a conversation marked as Escalated and review what happened before the handover.

Check:

• What the customer requested
• Which searches or tools Hugo used
• Whether information was missing
• Whether an instruction or routing rule required escalation
• Whether the customer explicitly requested a human
• Whether a tool failed before the escalation

An escalation is not always a failure. Hugo may be correctly following your configuration or handing over a request that requires human judgment.

Open the original conversation

Observability gives you the technical trace, but you may also need the complete support context from the Inbox.

Click Go to conversation from the selected trace.

The conversation opens in the Crisp Inbox, where you can:

• Read the full customer history
• Review private notes
• Check contact and account information
• See who handled the conversation
• Continue the investigation with your support team
• Reply to the customer when needed

Use Observability to understand the AI processing, then return to the Inbox when you need the broader customer context.

Export Observability traces

You can export Observability data when you need to investigate or share a conversation outside the view.

Export one conversation

Open the conversation you want to inspect, then click Export Trace.

This is useful when you need to:

• Share a specific problem with a developer
• Send a detailed example to Crisp support
• Keep a trace of an incident
• Compare the same issue before and after changing your setup

Export multiple traces

Click Export All from the conversation list to export the available traces.

Before sharing an exported trace, review whether it contains customer information, tool inputs, or private account data.

Observability vs Analytics

Observability and Analytics both help you improve Hugo, but they answer different questions.

Use Analytics to identify patterns, then use Observability to inspect specific conversations behind those patterns.

To learn more, read How to use Hugo Analytics to improve your AI Agent.

Best practices

Observability is most useful when you use traces to identify a specific improvement.

Start with failed and escalated conversations

These outcomes often provide the clearest opportunities to improve a tool, Workflow, instruction, or escalation rule.

Compare several conversations

Do not change your AI Agent setup based on one unusual conversation.

Review several examples to confirm that the issue is recurring before making broader changes.

Identify the failing layer

An unexpected answer does not always come from the AI model.

Check whether the issue comes from:

• Missing or outdated training resources
• An unclear Instruction
• A Routing or escalation rule
• A failed external tool
• An incorrect tool input
• An unavailable integration
• The selected AI model

Review inputs before blaming outputs

When a tool returns an incorrect result, check the parameters sent by Hugo and the data returned by the provider.

The final answer may reflect an incomplete or unexpected tool result.

Export traces before reporting bugs

An exported trace gives technical teams more context than a screenshot of the final customer answer alone.

Protect customer data

Traces can contain messages, customer details, tool parameters, authentication results, or returned account information.

Review exports before sharing them outside your authorized team.

Frequently Asked Questions

Still have questions which were not covered in this article? Here is a collection of the most frequently asked questions on this topic.

Can I inspect the tools Hugo called?

Yes. Tool call steps can display information such as the tool name, provider, input, output, execution time, logs, and status.

The exact details depend on the tool involved.

Can I see errors directly in Observability?

Yes. When an error occurs during the conversation, it can be displayed directly inside the relevant trace step.

Use the error message, tool input, and returned output to identify what needs to be corrected.

What should I do after finding an issue?

Identify which part of the setup caused the behavior before making changes.

Depending on the trace, you may need to update your training data, Instructions, Routing rules, model settings, Workflow, or connected tool.

Updated on: 21/07/2026