Introduction
Almost every Monitoring question is really a telemetry question. The screen shows what was sent; when it looks wrong, the cause is usually upstream of it.
Monitoring is empty, or nearly
Are any agents Active? — Monitoring can only show instrumented traffic. Agents in
Setup Pendingcontribute nothing, and unregistered agents contribute nothing at all.Is the time range wider than the traffic? — A quiet window over a genuinely busy agent is a range problem, not a data one.
Is the scope narrowed to an idle agent? — Reset to
All Agentsto separate "this agent is quiet" from "everything is quiet".
Note: An empty dashboard and a well-behaved estate look identical. This is the single most important thing to internalise about Monitoring. Before reporting that everything is healthy, confirm that something is actually being measured — the registry, not Monitoring, is where you check that.
The numbers do not match what I expect
Symptom | Usual cause |
Spend far below the provider invoice | Uninstrumented traffic, or models with no rate. Cost here is a floor |
Latency jumped, nothing deployed | A slower model or tool, or more calls per run — compare p99 on |
LLM calls rose, traces did not | Retries, a longer chain, or you enabled an LLM-based control or an AI judge |
Tool view is empty | Tool spans are not being emitted. Instrumenting the model client alone produces this exactly |
One agent dominates everything | Often correct. Confirm at agent scope before treating it as an anomaly |
Traffic dropped to zero
Take this seriously before assuming a reporting fault. In rough order of likelihood: the agent was decommissioned or its schedule changed; a deployment dropped the instrumentation; credentials or an ingest key rotated; or a network change blocked egress from where the agent runs.
The quickest split is to check whether the agent is doing its job at all. If it is working and not reporting, the problem is the exporter. If it is not working, Monitoring is telling you the truth.
Frequently asked
Does Monitoring show policy violations? — No. That is Incidents. Monitoring shows behaviour, Incidents shows breaches, Evaluation shows quality.
Can I get alerted when something changes? — Not from AI Governance today. There are no notifications on this surface — the review cadence is the mechanism, as it is for incidents. See Escalation & Notifications.
How far back does data go? — Telemetry ages out of retention after thirty days. Anything you need beyond that has to be captured while it exists.
Why do Monitoring and Cost Explorer show different totals? — Check the ranges and the scope first — a rolling window and a different period are not comparable. Both read the same telemetry.
An agent shows activity but no cost — Token counts are missing from the spans, or no rate exists for that model. Usage is observed independently of pricing.