Monitoring¶

Open an agent from the inventory or a platform to inspect its monitoring data. The available tabs depend on the connected platform and configured collection.

Agent overview¶

The Overview tab summarizes operational metrics, cost, and KPIs:

Operational: Reliability and activity, including interactions, average latency, errors, and token volume.
Cost: Model consumption and attributed cost over the selected time range.
KPI: The configured KPIs selected for display and their recent trends, when KPI data is available.

Operational and usage metrics¶

The Monitoring tab can include:

Uptime
Interaction and platform query volume
Interaction and platform error rates
Interaction and platform latency
Input, output, and total token volume
Daily, weekly, and monthly active users and conversations

Interaction metrics are generally derived from collected agent logs. Platform metrics are retrieved from the monitoring service exposed by the platform. The two sources can cover different requests or time ranges and should not be assumed to match exactly.

Usage data requires a stable user or conversation identifier in the collected data. Metrics remain unavailable when the platform does not expose the required information.

KPI metrics¶

The Monitoring tab displays the KPIs attached to the agent. Preferred direction indicates whether higher or lower values represent better performance. Configure alerts separately to detect KPI conditions that need attention.

Cost¶

The agent Monitoring tab attributes model calls, token usage, and cost to the selected agent. See Cost Management for portfolio and agent-level cost concepts.

Tools and agent properties¶

The Tools and Properties tabs describe the dependencies reported by the source platform. Structure can include tools, models, datasets, knowledge sources, and sub-agents. It helps identify shared dependencies and the downstream impact of a component change.

Structure is platform-declared metadata. The graph can be partial or unavailable when the connected platform does not expose enough information.

Agent Reviews¶

For supported agent types, the Testing tab associates Agent Reviews with an agent and displays their results. Use Setup to link an existing review or create one, and enable periodic reviews when needed. The platform’s Scheduling settings control the schedule; an agent can override the platform execution user.

Linked reviews run one at a time in their configured order. Removing a review from the list removes its association with the agent and preserves the review object.

Reviews assess agent behavior against defined checks. Their results complement operational metrics: an available agent can still fail a quality or safety check.

Dashboard controls¶

Monitoring pages provide controls that affect how metrics are calculated and displayed:

Time range selects the historical window, such as 24 hours, 7 days, 30 days, or 6 months.
Bucket selects the aggregation granularity, such as hour or day.
Timezone controls how bucket boundaries and timestamps are displayed.
Metrics selects the cards and series to display.

The latest point represents a completed metric bucket. It is not necessarily an instantaneous reading. Refreshing a metric requests updated source data but does not guarantee that the connected platform has published newer data.

Activity and collection health¶

The agent Activity tab shows collection operations scoped to the agent. Use it to distinguish three situations:

The platform does not support the metric.
Collection is configured but has not run yet.
Collection ran and failed or completed with warnings.

An agent can have usable historical metrics while a recent log or platform-metric fetch is failing. Review the operation status and error details before interpreting missing or stale charts.

For Dataiku agents, agent interaction logging describes the source interaction dataset used for deeper analysis and evaluation.