Introduction
Monitoring shows how your agents are behaving — trace volume, latency, error rates, token spend. For whether they are behaving within policy, see Incidents. For whether they are producing good output, see Evaluation.
Those three questions are genuinely separate, and this is the one place where they are easiest to confuse. An agent can be perfectly healthy by every measure on this screen and still be leaking customer data or answering wrongly.
Where to find it
AI Governance > Monitoring for every agent at once, or an individual agent's Monitoring tab for one.
Where Monitoring sits
Monitoring sits in the AI Observability group alongside Incidents. Monitoring shows how agents behave in aggregate; Incidents shows each time a control matched.
What Monitoring cannot do is stand in for Incidents or Evaluation. The three answer different questions:
Question | Surface | Signal |
Is it working? | Monitoring | Volume, latency, errors, spend |
Is it behaving? | Incidents | Controls that fired |
Is it any good? | Evaluation | Scores against a criterion |
What you can expect to find
Monitoring reads the same telemetry as the rest of the module — the traces and spans your instrumented agents emit — and aggregates it across time rather than showing individual runs. It has five tabs:
Tab | Shows |
| Trace count, error rate, latency (p50 / p99) |
| Call count, latency (p50 / p99), top 5 calls by model |
| Input and output tokens, total cost per day, cost per trace per day |
| Top 5 tools: run count, error rate, p50 latency |
| Top 5 run types, grouped by span type |
Pick an agent (default All Agents) and a range of Last 24 hours, 7, 14 or 30 days (default 30). Charts are bucketed by day in your time zone.
Monitoring can only show instrumented agents
Everything here is derived from traces. An agent that is registered but still Setup Pending contributes nothing, and an agent nobody registered contributes nothing at all.
An agent's Monitoring tab is disabled while it is Setup Pending. Models have no Monitoring tab.
This is the failure mode worth guarding against: a flat, healthy-looking Monitoring view is consistent with well-behaved agents and with agents that are not sending anything. Before reading anything into a quiet chart, check the registry for agents stuck in Setup Pending.
"An empty monitoring dashboard is not evidence of good behaviour. It is evidence of no telemetry, and those look identical from here."
How to read it, in order
Volume first, to establish what normal is — Every other number on this screen is only meaningful against a baseline. A doubling of errors on a day when volume tripled is an improvement.
Errors next, because they are unambiguous — An error is a run that failed. Unlike quality, it needs no judgement to interpret.
Latency third, at the tail rather than the average — Averages hide the runs users actually complain about. A stable mean with a growing tail is a real regression.
Spend last, against volume — Rising spend on rising volume is arithmetic. Rising spend on flat volume is a change in behaviour — a longer context, more retries, a switched model — and worth chasing. See How Cost & Tokens Are Calculated.
All agents, or one
Scope | Good for | Weak at |
All agents | Spotting that something changed | Saying which agent changed — a fleet aggregate absorbs a single agent's spike |
One agent | Confirming and diagnosing | Noticing in the first place, since you must already suspect that agent |
The workable habit is fleet-level for the periodic review, agent-level the moment anything looks wrong — then the agent's Observability tab for the individual runs behind it.