Agent Tracing Datapublic

Agent tracing captures detailed data for every interaction your agent receives. With this data, you can debug errors, monitor latency, track token usage, and understand the choices your agent makes as it builds a response. This reference explains every field you see when you inspect traces, either in real time from the Message Info panel in Agent Playground or later in the Observability log stream.

For instructions on how to enable traces and view their data, see How to Trace Agent Responses.

Agent Playground Metadata

Even without trace storage, you view the current trace in Agent Playground by clicking Message Info after you send a prompt. The panel shows token usage, latency, retrieved content, and other key metrics.

Each message shows Summary Metrics in Message Info. These metrics fall into two groups: Processing Time and Token Usage.

Processing Time

• Time submitted: when the agent receives the prompt.
• Time returned: when the agent finishes generating the response.
• Time to first token (TTFT): how long the agent waits before streaming the first token.
• Processing time: total time from submission to completion.

Use these values to track performance and isolate latency spikes.

Token Usage

• Prompt tokens: number of tokens in the user input.
• Response tokens: number of tokens in the agent’s reply.
• Total tokens: combined input-plus-output count.

These counts help you watch token-related costs and efficiency.

Input

The input field shows the exact prompt your user sends. For example:

“What’s the weather like in Austin tomorrow?”

Output

The output field shows the full response your agent returns. For example:

“Tomorrow in Austin will be mostly sunny with a high of 95 F.”

Knowledge-Base Retrieval

If the agent retrieves information from a knowledge base, the trace records:

• Knowledge base name: the knowledge base in use.
• File name: the data-source file that supplies context.
• Retrieved context: the text snippets the agent uses in its reply.

These details confirm what documents the agent is referencing to help generate its response.

Guardrail Triggers

If a safety guardrail fires, the trace lists:

• Guardrail name: the guardrail that triggers (for example, Jailbreak Detection).
• Reason triggered: why the guardrail activates, such as an unsafe prompt or instruction override.

You can use these fields to verify that safety policies work as intended.

Trace Metadata

To view full trace metadata, you must open the log stream from the agent’s Observability tab. The log stream opens in a separate window and shows stored sessions and traces in detail.

Each saved trace includes:

• Timestamp: when the platform receives and processes the request.
• Session ID: the unique ID for the conversation session.
• Trace ID: the unique ID for this trace.
• Status code: the HTTP status code the request received. 200 indicates success.
• Tool call ID: the ID of any tool call the trace makes.
• Has children: whether the trace spawns sub-calls (for example, a guardrail action).