Claude Platform adds native OpenTelemetry export to Databricks MLflow

Databricks MLflow Tracing is built on OpenTelemetry, so adding native export from the Claude Platform means Claude's traces can sit alongside traces from other instrumented services in the same MLflow experiment. Databricks receives the spans through Zerobus Ingest and stores them in Unity Catalog tables linked to that experiment, which lets teams inspect model calls and tool activity with the same debugging and evaluation tools they use for other services.

Once traces arrive, Databricks offers several ways to act on them. MLflow's Analyze trace action brings Genie Code into an investigation of a single execution, while MLflow Issues (available in Beta) uses Genie Code to group recurring problems across traces with links to supporting evidence. For quality judgments that need domain knowledge, Review Queues let experts assess selected traces and provide expected answers.

Traces also make agent behavior measurable: MLflow Agent Evaluation can run on Claude traces to test whether a change to prompts, skills, or tools improves the agent or causes a regression. The same approach supports checks such as whether an agent followed a required workflow or used appropriate tools, according to Databricks.