Local tracing¶

The Traces tab in the local experimentation dashboard lists every request the agent handled during a development session. Use it to inspect spans, tool calls, latency, token counts, and errors without deploying to the DataRobot platform.

For installation, GUI access, and port configuration, see the DataRobot experimentation plugin landing page.

Why use local tracing¶

Local tracing supports agent development in several ways:

• Faster iteration. Review traces, logs, and spans without leaving the local environment or breaking the development flow.
• Behavior verification. Validate agent behavior across multiple real requests before deployment, not only on a single demo input.
• Span-level debugging. Inspect individual spans to explain agent decisions instead of inferring them from the final response.
• Earlier issue detection. Catch errors, guardrail violations, and performance problems while changes are still local, before they reach production.

For guidance on adding custom spans and attributes that appear in local and deployed traces, see Implement tracing. After deployment, review traces in the DataRobot UI; see Tracing for custom applications.

Inspect OpenTelemetry traces¶

After installing the plugin, start the agent locally with task agent:dev. If the dashboard did not start automatically, run dr xp in another terminal. By default, it is available at http://localhost:8090. Configure the port with --port or DR_EXPERIMENT_PORT.

The dashboard lists traces. A trace is the record of one request to an agent, made up of spans for each step, such as LLM calls and tool calls.

The list isn't limited to your current session. It shows every trace exported to the entity you started dr xp for, from any agent or run. To find the ones you want, use the root span name and the filters.

Select a trace to see a detailed breakdown of its spans, including:

• Tool invocations and their execution order.
• Duration and latency for each span.
• Token counts, when reported by instrumentation.
• Associated error logs when a span or trace fails.

This is the same class of observability normally available only after deployment. With the experimentation plugin, it is available on the local machine from the start of development.

Filter and search traces¶

As request volume grows, use the dashboard controls to narrow results:

Analyze a trace¶

Trace analysis is a preview feature. To turn it on, start the dashboard with --enable-trace-analysis. The Tracing page then offers three ways to start an analysis:

• ✨ icon next to a trace's status badge: analyzes that trace.
• Analyze trace at the top of an open trace, next to the fullscreen and close icons: analyzes that trace.
• Analyze traces next to the filters above the trace list: analyzes every trace on the current page that has a recorded trajectory. This button is hidden while a trace is open.

A trace has a recorded trajectory only if its agent recorded its LLM calls with Tensile. Traces from an agent that isn't instrumented with Tensile, or that made no LLM calls, don't show the ✨ icon or the Analyze trace button. If no trace on the page has a recorded trajectory, Analyze traces is disabled.

Starting an analysis pulls the trajectory from DataRobot and opens an OpenCode session with the Tensile skills installed. A coding agent in that session reports how your agent performed and helps fix its mistakes. For setup and details, see Trace analysis.

Tracing troubleshooting¶

See also¶

• DataRobot experimentation plugin—install, configure, and open the GUI.
• Batch agent evaluation—score agent responses on a fixed dataset.
• Trace analysis—analyze recorded trajectories with Tensile and OpenCode (preview).
• Implement tracing—add custom spans and attributes.