Unify Logo Footer.svg
AI Governance
Logo
Agent Registry Troubleshooting & FAQ

Agent Registry Troubleshooting & FAQ

Logo

3 mins READ

Introduction

Nearly every registry question reduces to one of two: telemetry is not arriving, or something you expect to see is filtered out. Both have short diagnostic paths.

An agent is stuck on Setup Pending

Status waits on data, not on time. An agent leaves Setup Pending when its first trace arrives and not before, so the question is always whether anything is being sent.

  1. Has the agent run since you instrumented it? — An idle agent produces no traces and is indistinguishable from a broken one. Run it deliberately.

  2. Is the exporter configured where the agent actually runs? — Not on a developer machine. Exporters fail silently, so a misconfigured production deployment looks identical to an idle one.

  3. Does the key or endpoint belong to this record? — Traces sent with another agent's credentials arrive successfully — under that agent.

  4. Can the runtime reach the endpoint? — Egress rules and proxies are a frequent cause, and they do not surface as errors in the agent.

An agent is not in the list

Cause

Where it actually is

It was excluded

On its connection's detail page. The registry filters excluded assets out

It is a model

The Model Registry — the two lists are split by asset type

It is in another workspace

Switch workspace, or use the all-workspaces view if you have global access

You searched its description

Search matches the name only

It was never registered

Discovery found it; nobody acted on it

Telemetry columns are empty

Last trace, Activity, Cost, Incidents and Policies are computed from telemetry rather than stored on the record. Empty means one of three things:

  • They have not loaded yet. They resolve separately from the row and arrive a moment later.

  • There is no telemetry. Expected for anything in Setup Pending.

  • Nothing happened in the window. Activity is windowed, so a long-lived but currently idle agent reads as quiet.

Note: A zero in Cost alongside a healthy Activity is not a bug. Cost needs token counts on the spans and a rate for the model. Either can be missing while traces arrive perfectly well — see How Cost & Tokens Are Calculated.

A colleague sees a column I do not

Workspace hides itself outside the all-workspaces view and for users without global project read access. It is a permissions difference rather than a preference, and it is the usual explanation.

Frequently asked

  1. Why does Owner show someone who never claimed the agent? — Owner falls back to whoever created the record when no explicit owner was set. Treat inherited owners as unverified — see Ownership, Workspace & Department.

  2. Does Active mean the agent is governed? — No. Active means telemetry is arriving. Coverage is the Policies tab; enforcement is the connection tier. Three separate questions.

  3. Why is the Setup tab missing? — It renders only while the agent awaits its first traces, then disappears. That is intended.

  4. Can I change an agent's platform after registering? — The platform is chosen at registration. If it is wrong in a way that matters, register correctly and retire the old record — see Offboarding an Agent.

  5. Why do two records exist for one agent? — Usually a manual registration plus a discovery sweep finding the same thing. A discovered agent found on two connections is a single record, not a duplicate. Deleting one record disables it but keeps its history.

  6. Traces arrive but tool calls are missing — Only the model client is instrumented. Tool spans need the framework's own tracing, and without them no tool-surface control can fire.

  7. Can I bulk-register? — Register All on a connection's discovered banner takes the whole outstanding set. There is no bulk path for manual registration.