You can now bring Snowflake Intelligence (Cortex Agents) under Monte Carlo's Agent Observability coverage!

From Settings > Agent Observability, connect a native Cortex Agent by selecting a Snowflake warehouse and choosing from a list of discovered agents -- no custom instrumentation or code changes required.

Once connected, Cortex trace and span data flows in automatically, powering the full Agent Observability experience: trace viewer, conversation viewer, trace dashboards, and all four agent monitor types.

Learn more and get started here: https://docs.getmontecarlo.com/docs/snowflake-intelligence-cortex-agents

Metric monitors now support out-of-the-box ML metrics. Select from prebuilt, industry-standard regression metrics — RMSE, MAE, MAPE, R-Squared, and Mean Error — to measure how far off a model's predictions are, plus classification accuracy to track how often a model gets the right answer.

Just pick the prediction and actual columns, and monitoring is configured in clicks.

Coming soon: out-of-the-box drift detection metrics including PSI, KS Test, and JS Divergence for automatic distribution shift detection

Learn more here: https://docs.getmontecarlo.com/docs/available-metrics#ml-metrics

The Asset Summary tab now surfaces all monitor results and trends in a single view.

Metric monitors, custom SQL, validations, and comparisons appear alongside existing pipeline health signals, each with a time-series chart that respects the global time window.

Inline controls let you switch fields, segments, and variables without leaving the page.

Toggle between a results view and a Monitor list view, and a customization option allows you to choose which monitors to show or hide.

Multi-turn agent interactions can now be viewed as a single, chronological conversation in Trace Explorer. Traces sharing a conversation ID are automatically grouped, displaying user messages and agent responses in order with timestamps, role labels, and turn numbers.

Each turn links directly to its underlying trace, and clicking any message opens a detail panel with the full content. This makes it easy to follow the complete back-and-forth of an agent conversation in one place and drill into any individual trace when you need more detail.

You can now track a metric over time without triggering alerts.

What's New:

1. Track metrics before turning on alerts Metric, Custom SQL, and Comparison monitors now support a Track mode. When defining alert conditions, you can choose "Track" to run the monitor and collect the metric over time without setting thresholds or triggering notifications. NOTE: Save the monitor in “Enabled” status (not as a draft) to start it running.

2. Track metrics on your custom dashboard You can now add any chart to a custom dashboard using the dashboard icon on the chart (top right). This lets you keep an eye on the signals you care about in one place. Charts can be added from:
   1. Monitor results
   2. Asset results
   3. Data profiling results (Yes, now you can observe profiling metrics over time)

Why this matters

In some cases you want to observe a signal (null %, row count, any custom metric) before deciding what should actually alert. Track mode lets you build confidence in the data before operationalizing it.

With Track mode and dashboards you can:

• Observe metrics before setting thresholds
• Validate assumptions without creating alert noise
• Keep a persistent place to check data health
• Turn tracked signals into alerts when you’re ready

The troubleshooting agent now presents supporting evidence in a structured timeline format, including PR diffs, making it easier to follow the reasoning behind each recommendation.

You can also point to any specific piece of evidence and ask the agent to reconsider it, with the option to include additional context or clarification. This gives you more control over the troubleshooting process and helps you arrive at the answer you need faster.

Agent evaluation monitors and metric monitor prompt configurations now support custom LLM model names.

You can type any model name directly into the model selector instead of being limited to a predefined list, making it easy to use the latest models as soon as they are available.

The dropdown still shows all predefined options for convenience, with custom values clearly labeled. This gives your team the flexibility to route LLM requests to any model identifier your environment requires.

You can now create production-ready LLM-as-a-judge evaluations by simply describing what you want to measure. Type a short description of the dimension you care about, hit Generate, and get a complete eval prompt ready for production.

Starter templates are included for common evaluation dimensions like answer relevance, helpfulness, task completion, language match, clarity, prompt adherence, and semantic similarity. Advanced controls let you fine-tune scoring criteria and strictness levels to match your specific requirements.

You can now track and manage individual breached rows from Custom SQL and Validation monitors directly in Monte Carlo. When a primary key is configured on a monitor, breached rows are tracked across runs in a new Exceptions tab. From there, you can assign an owner, set a resolution status, add comments, take bulk actions, and track how long each exception has been open.

Learn more here: https://docs.getmontecarlo.com/docs/exception-management

Agent assets now include out-of-the-box dashboards showing trace volume, latency distributions across P50/P95/P99, token consumption trends, and error rates. All with automatic period-over-period comparisons. No configuration is required; connect your OpenTelemetry traces and the views are ready.

Whether you need to spot spikes in token usage, catch latency drifting upward, or confirm your agents are behaving as expected, these dashboards give your team immediate visibility from day one with a natural path to production-grade alerting as your agents mature.