Why is an agent's COST empty?

Cost is derived from trace data. An agent with no traces has no cost to report.

Can two agents have the same name?

Yes — names come from the source platforms and are not forced to be unique. Use PLATFORM and OWNER to disambiguate.

Does removing an agent from the registry stop it running?

No. The registry governs agents; it does not operate them. Removing a row removes your visibility, not the agent.

Why does the count change between visits?

Connected platforms are swept continuously, so newly created agents appear on their own.

Introduction

The Agent Registry is the list of every AI agent your organisation is accountable for. Discovery finds agents; the registry is where they live, where they get an owner, and where policy reaches them.

If someone asks "how many AI agents do we run, who owns them, and what have they cost this month", this is the page that answers.

Where to find it

AI Governance > Agent Registry

Access to AI Governance is granted through platform roles. Ask your workspace administrator, or see Settings > Governance > Roles for how permissions are assigned.

How agents get into the registry

Two intake routes, and the product states them in one line on its own empty state: agents are synced from discovery or added manually.

Discovered, then registered — Connect a platform and Discovery sweeps it. What comes back is a list of discovered assets on that connection — not yet registry entries. Open the connection to see them, then register the ones you want to govern with Register All, or one at a time from the Register column. Turn on Auto-register agents in the connection's settings and newly discovered agents are added for you.
Registered by hand — + Register Agent opens a form titled Enter manually. Use it for agents on platforms Discovery cannot sweep, and for agents built directly on a framework such as LangChain. See Register an Agent Manually.

Refer: Discovering an asset and registering it are two separate acts, and only the second one brings it under governance. See Working with Discovered Agents.

"Sweep first, then register deliberately. Auto-register is the right setting once you know what a platform holds — rarely before."

The registry table

One row per agent. The count beside the page title is the number of agents currently registered.

Column	What it tells you
`AGENT NAME`	The agent's name. Opens the agent's detail view.
`PLATFORM`	Where the agent runs.
`STATUS`	`Active` or `Setup Pending`.
`LAST TRACE`	When the agent last ran, as a relative time. A dash means no trace has been received.
`ACTIVITY`	Volume, counted in traces — for example `41 traces`.
`COST`	What the agent has spent.
`INCIDENTS`	Governance incidents raised on this agent.
`POLICIES`	How many policies apply to it. `0` means the agent is inventoried but not governed.
`WORKSPACE`	The workspace the agent belongs to.
`OWNER`	Who is accountable for it.

Two further columns are hidden by default; the hidden-columns control above the table shows how many. The table also carries Filter, Sort and Search.

Traces are the unit of activity

An agent's activity is measured in traces. One trace is one end-to-end run of the agent. This matters when you compare the Agent Registry with the Model Registry, which is the same table with one substitution: models are measured in spans, the units of work inside a trace.

Reading 0 traces on an Active agent is common and usually correct — the agent exists and is configured, and has not run since it came under governance.

The three statuses

Active — Registered, receiving traces, and ready to be governed.
Setup Pending — Registered, but no traces have arrived yet. The agent's Setup tab is where you fix that — it shows the SDK setup steps for the agent's platform and Waiting for traces… until the first trace lands. The status clears on its own once it does.
Excluded — Discovered but not registered — visible on the connection that found it, and not under governance. This is the default state of everything a sweep returns until you register it.

Inventoried is not governed

The single most useful habit with this table is reading the POLICIES column rather than the row count.

An agent in the registry with 0 policies is one you have found and named. Nothing checks what it does. The gap between total agents and agents with at least one policy is the real measure of programme coverage, and it is the number worth reporting.

Refer: To bring agents under policy, see Apply a Policy to Assets.

What an agent row opens

Clicking an agent opens its detail view — seven tabs: Overview (Setup while the agent is pending), Observability, Monitoring, Outcomes, Evaluation, Incidents, Policies.

Best Practices

Report coverage, not count — "116 agents registered" is an inventory. "41 of 116 have a policy applied" is a programme status.
Give every agent a named human owner — Incidents need somebody to route to. An unowned agent produces incidents nobody triages.
Work the Setup Pending queue on a cadence — Each one is an agent you registered that is not yet reporting. Its Setup tab tells you what it is waiting for.
Sort by LAST TRACE to find abandoned agents — An agent nobody has run in months still holds permissions and still carries risk. Retiring it is the cheapest risk reduction available.
Use WORKSPACE and OWNER from day one — Retrofitting ownership across a few hundred rows is a project. Setting it at registration is a field.

Agent Registry Overview