Monitor AI Agents

On This Page

Monitoring helps you understand how AI agents are being used, identify performance issues, and spot AI agents that may need instruction or tool refinements.

For testing guidance, see Create and Configure an AI Agent.

Use the Monitor tab

The Monitor tab tracks how your AI agents are performing over time. The run history uses a timeline layout that displays each step chronologically—including tool calls, AI agent reasoning, and outputs—so you can follow the AI agent's execution path from start to finish.

Note: You can stop active AI agent runs clicking the Stop button. This is useful for canceling AI agents that are consuming excessive resources or producing incorrect results.

You can monitor your AI agents from two places:

The Monitor tab in the main navigation panel. This screen displays a log of all AI agent requests, including their current status and detailed metadata. Use this view to track the progress of each request, identify issues, and review historical activity. From this screen, you can click a Request ID to view detailed information about the AI agent's processing steps and outputs.

Scroll or drag to pan · ⌘+/⌘− to zoom · Esc to close
The AI agent-specific Monitor tab. To view request activity for a specific AI agent, open the AI agent and go to its Monitor tab. This shows only the requests handled by that AI agent.

Scroll or drag to pan · ⌘+/⌘− to zoom · Esc to close

You can also view all AI agent runs across all AI agents in the AI Agent Activity tab of the Monitor view.

Monitor chat agents

Chat agents appear in the agent-specific Monitor tab. When a chat agent is deployed via the agent chat component, end-user interactions generate runs that are visible in the chat agent's Monitor tab.

You can also view all AI agent runs across all AI agents in the AI Agent Activity tab of the Monitor view.

AI agent summary table

Each row in the table shows key metrics for an AI agent:

Field	Description
Run ID	A unique identifier for each request. Click to view the full request details.
Version	The version number of the AI agent that processed the request.
Status	The current state of the request.
Agent Name	The name of the AI agent that handled the request.
Request	A description of the request.
Token Usage	The number of tokens consumed during the AI agent's processing.
Tool Calls	The number of tools the AI agent used while processing the request.
Initiator	The person who initiated the request.
Duration	The time the AI agent took to complete the request.
Start Time	The date and time when the AI agent request began processing.

MCP tool calls in the Monitor tab

When an AI agent calls an MCP tool, the tool call appears in the step history alongside native tool calls. Each MCP tool call shows the tool name, the request sent to the MCP server, and the response received.

MCP tool calls are counted in the Tool Calls metric for each run. They contribute to the same maximum tool call limit as native tools.

If an MCP tool call fails, the error message from the MCP server appears in the step details. Common failure reasons include authentication errors, server unavailability, and vendor-imposed rate limits.

Developer feedback in the Monitor tab

Developers with Viewer permissions or higher can provide feedback on their own AI agent runs directly from the Monitor tab. Click a Run ID to open the run detail view, then use the thumbs-up or thumbs-down control to rate the overall run outcome.

Feedback submitted from the Monitor tab follows the same model as feedback in the Evaluate tab. Ratings contribute to evaluation metrics when the run is associated with a test case. For the full feedback workflow and evaluation metrics, see Evaluate AI Agents.

Parallel tool call display in run details

When an AI agent executes parallel tool calls, the run detail view groups those calls under a single step:

Step grouping: All parallel tool calls share the same step ID. The run detail view displays them as a grouped set to indicate concurrent execution.
Duration reporting: The duration for a parallel step reflects the wall-clock time from dispatch to completion of the last tool call, not the sum of individual call durations.

For parallel execution constraints, see Parallel tool execution.

Feedback

Was this page helpful?