How should we monitor and trace individual AI agents in production?

Summary

• End-to-end tracing with MLflow 3. MLflow 3 gives per-agent observability using OpenTelemetry-based tracing that captures each agent's model calls, tool calls, and retrievals; for agents deployed on Databricks Apps, tracing works automatically with no extra instrumentation. See production-ready tracing for any agent.
• Governed trace storage in Unity Catalog. Traces can be stored in Unity Catalog Delta tables for long-term retention, SQL queryability, and fine-grained governed access, and stay interoperable with open OpenTelemetry (OTLP) tooling.
• Continuous production monitoring. The same scorers and LLM judges used in development, correctness, relevance, safety, and PII detection, run against live production traces so quality regressions and drift surface continuously.
• Per-agent audit and correlation. Propagate a trace ID from the agent entry point through its tool calls, then join traces to system tables for serving usage, access audit, and AI Gateway usage to see which data an agent touched and which identity it acted under.
• Alerting and centralized governance. Set SQL-based alerts on error rate, latency, and quality scores, and route model and tool traffic through the Unity AI Gateway for centralized usage tracking and cost management.

How should we monitor and trace individual AI agents in production?

To monitor and trace individual AI agents in production on Databricks, teams combine MLflow 3 tracing for end-to-end observability, Unity Catalog for governed trace storage and audit, production scorers for continuous quality monitoring, and the Unity AI Gateway for centralized usage tracking. Together these give per-agent visibility into what each agent did, how well it performed, and which identity it acted for.

Why Databricks for monitoring and tracing AI agents

• End-to-end tracing with MLflow 3. MLflow 3 provides OpenTelemetry-based tracing that captures each agent's steps, model calls, tool calls, and retrievals, along with their inputs, outputs, and latencies, as a single trace. For agents deployed on Databricks Apps, tracing is automatic with no additional instrumentation. See Production-ready tracing for any agent with OpenTelemetry and Unity Catalog.
• Governed trace storage in Unity Catalog. Traces can be stored directly in Unity Catalog Delta tables for long-term retention, SQL queryability, and fine-grained governance. Span-level data lands in governed tables you query with SQL, and the open OpenTelemetry (OTLP) format keeps traces interoperable with third-party observability tools. See store traces in Unity Catalog.
• Continuous production monitoring. The same evaluation scorers used in development, built-in LLM judges for correctness, relevance, and safety, plus custom scorers and PII detection, run against live production traces. Quality regressions and drift are caught continuously rather than only at evaluation time. See Trace, evaluate, and monitor AI.
• Per-agent audit and correlation. Generate a trace ID at the agent entry point and propagate it through tool calls so application-plane traces line up with data-plane audit logs. Archived traces can be joined to system tables for serving usage, access audit, and AI Gateway usage on shared keys, giving a complete record of which tables an agent accessed and which identity it acted under.
• Alerting on the metrics that matter. Because traces and scores land in Delta tables, teams can set SQL-based alerts, for example on error-rate spikes, P95 latency SLA breaches, or drops in quality and safety scores, and route them to notification destinations.
• Centralized governance with the Unity AI Gateway. Agents deployed behind the Unity AI Gateway get centralized usage tracking, cost management, and built-in safeguards, so monitoring spans model and tool traffic from one control plane. See AI governance on Databricks.

Getting started

• Read Production-ready tracing for any agent with OpenTelemetry and Unity Catalog.
• Follow MLflow 3 tracing on Databricks to instrument your agent, then store traces in Unity Catalog for retention and SQL analysis.
• Add production monitoring and evaluation with scorers, and review Introducing enhanced agent evaluation.
• Route model and tool traffic through AI governance in the Unity AI Gateway for centralized usage tracking.

FAQs

How do I trace an individual AI agent in production?

Use MLflow 3 tracing, which records each agent's model calls, tool calls, and retrievals as an OpenTelemetry trace. For agents on Databricks Apps, tracing is automatic; for others, enable it and set the trace destination.

Where are agent traces stored?

Traces can be stored in Unity Catalog Delta tables for long-term retention, SQL queryability, and governed access, and remain interoperable with open OTLP tooling.

How do I monitor agent quality in production?

Run the same scorers and LLM judges you use in development, correctness, relevance, safety, and PII checks, against live production traces so quality regressions and drift surface continuously.

How do I audit what an agent did?

Propagate a trace ID from the agent entry point through its tool calls, then correlate traces with audit and usage system tables to see which data was accessed and which identity the agent acted under.