Introduction

A model record and an agent record are the same underlying object with a different asset type, so most of what you know about one carries over. What does not carry over is the interesting part: a model is measured in spans rather than traces, and it has no setup step at all.

Where to find it

AI Governance > Model Registry > open a model.

The list columns

Column	Shows
`Model name`	The model as the serving platform names it
`Platform`	The connected application serving it
`Status`	`Active` or `Inactive`
`Last span`	When this model was last called
`Activity`	Span volume in the window
`Cost`	Spend attributed to this model
`Incidents`	Incidents raised on it
`Policies`	Policies covering it
`Workspace`	The project the record lives in
`Owner`	The accountable person
`Last updated`	When the record last changed

Every difference from an agent record, in one table

	Agent	Model
Unit of activity	Traces	Spans
Recency column	`Last trace`	`Last span`
Statuses	`Active` / `Setup Pending` / `Excluded`	`Active` / `Inactive`
Detail tabs	Seven	Five
How it is added	Registered — you supply the identity	Imported — you pick from what the connection returns
Instrumentation	An SDK or OTLP endpoint you wire up	None. Spans arrive through whatever calls it

Why spans, not traces

A trace is one run of an agent: a sequence of steps with a beginning and an end. A model does not have runs — it has calls. Each call is a span, and the same model may appear in spans belonging to many different agents' traces.

This changes what the activity number means. An agent's trace count is how often that agent ran. A model's span count is how often anything called it, across every agent and every workflow in scope. A model with high activity and no obvious owner is usually shared infrastructure, and that is the most useful thing the number tells you.

Note: A model's cost and an agent's cost are not additive. The spend on a model span is also spend inside some agent's trace. Adding a model's cost to an agent's cost double-counts the same call. See How Cost & Tokens Are Calculated.

Active and Inactive

A model has two statuses, not three. There is no Setup Pending, because there is nothing to set up — you cannot instrument a model the way you instrument an agent. Either spans are arriving through the connection or they are not.

There is also no Excluded. Exclusion is a decision made on a discovered asset at the connection, and it applies to models there just as it does to agents; it simply is not a status this registry renders.

The five tabs

`Overview`	Identity, Requests tiles, activity chart, incident summary
`Observability`	The model's spans
`Evaluation`	Span-level evaluators scoring this model's calls. `Evaluation Settings` opens a drawer with one setting: how often spans are swept for evaluation
`Incidents`	Incidents raised on this model
`Policies`	Policies currently covering it

Three tabs an agent has are absent: Setup (nothing to instrument), Monitoring, and Outcomes (an outcome is credited to an agent, which is the thing that accomplishes work — a model contributes to outcomes it is never credited with).

The model Overview counts Requests — calls to the model — rather than runs. The model Evaluation tab runs span-level evaluators, and its Evaluation Settings drawer has a single setting: how often spans are swept for evaluation.

Identity fields

Owner, Workspace and Department are set when the model arrives — an imported model takes the person who imported it as owner — and cannot be edited from the model page afterwards. They matter more rather than less: a shared model with no named owner is a governance gap that spreads across every agent calling it. See Ownership, Workspace & Department.

"An ungoverned agent is one team's problem. An ungoverned shared model is everyone's."

When a model shows no spans

The empty state says it plainly: spans appear once the model starts processing requests and generating execution spans. In practice there are three readings, and they are worth separating before investigating:

Nothing has called it — An imported model that no agent uses. Common, and not a fault — it is inventory.
Something calls it outside instrumented paths — Traffic that never passes through a traced agent produces no spans here. The model is in use and invisible.
The calling agent is not sending traces — An agent still in Setup Pending produces no spans for the models it calls either.

Model Record Reference