Introduction
Monitoring opens on every agent at once, and narrows to one. Those two scopes are good at opposite things, and using the wrong one is why a real problem can sit unnoticed on a dashboard somebody checks every morning.
Where to find it
The scope picker sits above the switchers, defaulting to All Agents. The time range sits opposite it: Last 24 hours, 7, 14 or 30 days (default 30), bucketed by day in your time zone.
What each scope is good and bad at
All agents | One agent | |
Good at | Noticing that something changed | Explaining what changed |
Bad at | Saying which agent | Noticing in the first place |
Fails by | Absorbing a single agent's problem into the aggregate | Only ever showing you what you already suspected |
Use for | The periodic review | The investigation |
Note: The fleet view has a blind spot proportional to your fleet. One agent doubling its error rate is obvious across five agents and invisible across fifty. The more successful your AI programme is, the less the aggregate can tell you — so the review habit has to change as the estate grows.
The workflow that works
Start at fleet scope, at a fixed cadence — You are looking for a change in shape, not a number. Anything unexplained becomes step two.
Check what kind of step moved —
Run Types,LLM CallsandToolsshow whether model calls or tool calls account for the change. See Run Types.Narrow to candidate agents — The registry table, sorted by
IncidentsorActivity, is often faster than stepping through the picker.Confirm at agent scope — Same switchers, one agent. This is where a suspicion becomes a finding.
Drop into individual runs — The agent's
Observabilitytab — see Reading a Trace.
Compensating for the blind spot
Because the aggregate degrades as the estate grows, fleet-level review needs a supplement rather than more attention. Three that work:
Review the registry table, not only Monitoring. Per-agent incident and activity counts sit side by side there, so an outlier is visible as a row rather than hidden in a total.
Let incidents do the noticing. A control that fires is agent-specific by construction — it names the asset. That is a per-agent signal arriving without anyone scoping a view.
Review the newest agents specifically. A recently registered agent has the least established baseline and contributes least to the aggregate, which is exactly the combination the fleet view handles worst.
"The fleet dashboard gets less useful with every agent you add to it. Plan the review for the estate you will have, not the one you started with."
The time range is part of the scope
Range choice changes what is visible as much as agent scope does. A window wide enough to establish a baseline will flatten a spike; one narrow enough to show the spike has no baseline to judge it against.
The practical habit is to look at both: a wide range to know what normal is, a narrow one to see what just happened. A finding that survives both is real.