Introduction
Three nouns carry all of the telemetry in AI Governance: session, trace, span. They nest, and almost every other number in the product is derived from them — activity, cost, latency, evaluation coverage, and the evidence attached to an incident.
Getting these three straight makes the rest of the product legible.
Where to find it
AI Governance > Agent Registry > select an agent > Observability
The three levels
Level | What it is | Roughly |
| A related group of traces, held together by a | One conversation |
| One end-to-end run of the agent. | One turn |
| One unit of work inside a trace. | One model call or tool call |
A user asks three questions in one conversation: that is one session containing three traces. Each trace may call a model twice and a tool once: three spans per trace. Nine spans, three traces, one session — from one interaction.
"Sessions are what the user experienced. Traces are what the agent did. Spans are what it cost."
Traces versus spans
The Agent Registry measures activity in traces; the Model Registry measures it in spans. This is not an inconsistency — it is the same data read at the level each asset operates.
An agent's unit of work is a complete run. A model does not run agents; it serves calls, and a call is a span. When the same underlying activity is attributed to a model rather than an agent, the natural unit changes with it.
Practical consequence: never compare an agent's trace count with a model's span count. They are different units, and the span number will always look larger.
What a session record holds
Field | Description |
| The identifier grouping traces into one session. |
| What opened the session. The most useful column for finding a session again. |
| Total cost of every trace in the session. |
| Tokens consumed across the session. |
| How many runs the session contains. |
| Wall-clock length of the session. |
| When it opened. |
Session-level tiles summarise the same fields across the window: Total Sessions, Avg Traces/Session, Avg Session Time, Avg Tokens Used.
Avg Traces/Session is the quiet diagnostic here. A number climbing over time usually means users are asking again because the first answer did not land.
The All / Evaluated toggle
Every level carries a toggle between All records and Evaluated ones.
Evaluation is normally sampled rather than exhaustive, because scoring costs money. So Evaluated is a subset, and the ratio between the two is what the COVERAGE tile on the Evaluation tab reports as records scored out of records sampled.
When you cite a quality score, cite the coverage with it. A blended score over two per cent of traffic is a different claim from the same score over all of it.
What is derived from this data
Activity and LAST TRACE — Straight from trace records.
Cost — Aggregated from spans, rolled up to traces, sessions and agents. Every cost figure in AI Governance ultimately comes from here — including the cost of running governance itself. See Where Spend Comes From.
Latency and error rate — Trace-level, surfaced as p50, p90 and p99 on
Monitoring.Incident evidence — An incident points at the trace that produced it, which is what makes investigation a click rather than a correlation exercise.
Evaluation scores — Evaluators score at
SessionorTracescope — one reason the distinction matters operationally.
Known limitations
Records are agent-scoped — There is no global trace explorer. You reach traces through an agent.
Sessions depend on the source emitting a case identifier — Where a platform does not group runs, traces may appear without meaningful session grouping.
No traces, nothing downstream — An agent that has never run has no cost, no latency, no evaluation coverage and no incidents. That is data absence, not a fault.
Captured detail varies by provider — What lands in a span depends on what the source platform emits.