How do I monitor custom AI agents in production?

Summary

• Monitor custom AI agents on Databricks with MLflow 3, which unifies tracing, evaluation, and observability across the agent lifecycle so the same quality checks you use in development also run on live production traffic.
• MLflow Tracing records the inputs, outputs, and metadata of every intermediate step of an agent request, follows the OpenTelemetry (OTel) format, and can be enabled with autologging for popular agent frameworks.
• Production monitoring applies built-in and custom LLM judges and code-based scorers — for safety, relevance, correctness, and retrieval quality — to production traces, with configurable trace sampling to control evaluation cost.
• Agent Evaluation measures the quality, cost, and latency of agentic applications, including latency percentiles, error rates, throughput, and token-usage and cost patterns over time.
• Collect human feedback through the Review App and labeling sessions in development, and annotate production traces with user feedback via APIs for continuous improvement.

How to monitor custom AI agents in production

A custom AI agent chains model calls, tool invocations, and retrieval steps, so monitoring it means seeing what happened inside each request, scoring the quality of the results, and watching cost and latency on live traffic. On Databricks this runs on MLflow 3, which unifies tracing, evaluation, and observability across the agent lifecycle — so the judges and scorers you define during development are the same ones that measure quality in production.

Why Databricks monitors custom AI agents with MLflow 3

• End-to-end tracing. MLflow Tracing records the inputs, outputs, and metadata for each intermediate step of an agent request, giving full visibility into how the agent reached its answer. Traces follow the OpenTelemetry (OTel) format and can be enabled with a single line of autologging, with automatic instrumentation for popular agent frameworks.
• Deploy-anywhere observability. Agents running on AWS, GCP, on-premises, or Kubernetes can connect to MLflow 3 for real-time observability, so you monitor agents wherever they run.
• Built-in and custom scorers. MLflow 3 includes built-in LLM judges for safety, relevance, correctness, and retrieval quality, plus support for custom LLM judges and code-based scorers tailored to your business requirements. It also integrates third-party evaluation libraries as scorers.
• Production monitoring on live traffic. Production monitoring applies the same judges and scorers used in development-time evaluation to production traces, enabling continuous quality measurement on live traffic. You can configure trace sampling rates to control how often evaluation runs and manage its cost.
• Quality, cost, and latency together. Agent Evaluation measures the quality, cost, and latency of agentic applications — including latency percentiles (such as P50/P95/P99), error rates and throughput over time, and token-usage and cost patterns you can compare across time periods.
• Conversational trace analysis. A natural-language interface lets you inspect traces, evaluation runs, and scorers conversationally within your MLflow experiment — for debugging a specific trace, reviewing performance metrics, or assessing quality — without writing queries by hand.
• Human feedback in the loop. During development the Review App supports quick checks through a chat UI and structured expert feedback through labeling sessions; in production, APIs let you annotate traces with end-user feedback to drive continuous improvement.

Getting started

• Start with MLflow 3 for generative AI to enable tracing and observability for your agent.
• Follow trace, evaluate, and monitor AI to add built-in and custom scorers and turn on production monitoring against live traces.
• Use evaluate, debug, and optimize agents to measure quality, cost, and latency for your custom agent.
• Read the MLflow 3.0 announcement and enhanced agent evaluation for what tracing and evaluation unlock.

FAQs

How do I see what a custom AI agent did on each request?

Enable MLflow Tracing, which records the inputs, outputs, and metadata for every intermediate step of an agent request in the OpenTelemetry format. Autologging turns this on with minimal code, and traces are viewable in your MLflow experiment.

Can I run the same quality checks in production as in development?

Yes. MLflow 3 production monitoring applies the same LLM judges and scorers you use during development-time evaluation to live production traces, so quality is measured continuously. Trace sampling rates let you control evaluation frequency and cost.

How do I track an agent's cost and latency?

Agent Evaluation measures quality, cost, and latency together, including latency percentiles, error rates, throughput, and token-usage and cost patterns over time, so you can spot regressions and compare performance across periods.

How do I capture human feedback on agent responses?

In development, the Review App enables quick checks and expert labeling sessions. In production, APIs let you annotate traces with end-user feedback, which feeds back into evaluation and continuous improvement.