Introduction

A trace is one run of an agent; an LLM call is one request to a model inside it. The ratio between the two is one of the most diagnostic numbers in the product, and it is only visible because these are separate views.

Where to find it

AI Governance > Monitoring > LLM Calls.

The tab shows call count, latency (p50 / p99) and the top 5 calls by model, per day.

Calls per trace is the number to watch

One run can make one model call or twenty. How many it makes is a property of how the agent is built, and a change in that ratio is a change in behaviour — usually one nobody intended.

Traces	LLM calls	Reading
Flat	Flat	Steady state
Up	Up proportionally	More usage. Expected
Flat	Up	Each run is doing more work — retries, a longer chain, or a loop
Down	Flat or up	Fewer, much heavier runs. Usually a failing path retrying

The third row is the one worth a standing eye on. It is the earliest visible sign of a regression that will show up as a cost problem weeks later, and by then the cause is harder to find.

Note: An LLM call is not always the agent thinking. Governance controls that use an LLM check make model calls too, and so does an AI judge scoring a session. Turning on either adds calls without the agent doing anything differently — so correlate a step change with your own configuration changes before treating it as agent behaviour.

What this view answers

How much model work is happening — Independent of how many runs it took.
Which models are called most — The top 5 by model. A change in mix is often a deployment nobody flagged as significant.
Where latency actually comes from — Model calls are usually the bulk of a run's duration. An agent that feels slow is often making more calls rather than slower ones.
What it does not show — LLM Calls has no failure metric. Use the Traces tab for error rate.

Watching the model mix

Model choice is usually made once, early, and then inherited by everything built afterwards. Two patterns are worth catching:

A default model used everywhere. Efficient to reason about and a single point of cost. A change here moves every agent at once — which is a risk and an opportunity.
The largest model used for trivial steps. Classification, routing and extraction rarely need the most capable model. This is the most reliable cost saving available and it costs no quality when scoped to the right steps.

"Most AI overspend is not too many requests. It is the right number of requests going to a model that was chosen once and never revisited."

This view is not the cost view

Call volume and spend move together but are not the same. A run making fewer calls with much larger contexts can cost more than one making many small calls, because input tokens dominate.

Read this view for behaviour and the Cost & Tokens switcher for money — see Cost & Tokens.

LLM Calls