Introduction
Opening an agent gives you seven tabs. They are not seven views of the same data — each answers a different question, and knowing which one to open saves the tour.
One structural fact worth stating up front: observability in UnifyApps is agent-scoped. There is no global trace explorer across every agent. You start from an agent and work down. That is a deliberate consequence of the registry being the spine of the product.
Where to find it
AI Governance > Agent Registry > select an agent
The tabs
Tab | The question it answers |
| Is this agent healthy, what does it cost, and who owns it? |
| Why has this agent sent no traces, and how do I fix it? Replaces |
| What exactly did it do? Individual sessions, traces and spans. |
| How is it behaving in aggregate, over time? |
| What business value has it produced? |
| How good are its responses, and what does scoring them cost? |
| What has it done wrong? |
| What rules apply to it? |
Overview
The landing tab, and the one to read first when an agent is unfamiliar.
Agent Activity carries four tiles over a period picker — Last 7, 14, 28 or 90 days, defaulting to Last 7 days:
Trace count — How many end-to-end runs in the window.
Total cost — What those runs cost.
Error rate — Share of traces that ended in error.
Latency P90 — The slow-tail experience, not the average.
Below the tiles, a Traces bar chart shows volume by day across the window — the fastest way to spot an agent that stopped, or one that suddenly started.
On the right, Agent Details holds the descriptive record: description, Connection type, Status, Owner, Workspace and Department. Below it, an Incidents block headed "Governance incidents raised on this agent" with Total incidents, Critical, High, Medium and Low, a View incidents ↗ link, and a table of the most recent incidents. Empty states read "No incidents" and, for policies, "No policies yet" with a Set up policies link.
Setup
The tab that appears when an agent is registered but has not reported yet — the same condition that shows Setup Pending in the registry. It shows the SDK setup steps for the agent's platform.
The tab shows Waiting for traces… until the first one lands, then the agent turns Active on its own. Where a platform has no instructions yet, it reads No setup available.
Observability
The record explorer — where you go when you need the specific thing that happened rather than a summary.
Sub-switchers move between Sessions, Traces and Spans, which are the three levels of the same data. A second toggle, All and Evaluated, filters to records an evaluator has scored.
Session-level tiles: Total Sessions, Avg Traces/Session, Avg Session Time, Avg Tokens Used. Session columns: CASE ID, FIRST MESSAGE, COST, TOKENS, TRACES, DURATION, START TIME.
Refer: For how sessions, traces and spans relate, read Sessions, Traces & Spans.
Monitoring
The aggregate charts. Where Observability shows you one run, Monitoring shows you the shape of thousands.
Sub-switchers: Traces, LLM Calls, Cost & Tokens, Tools, Run Types. Charts include trace count split by success and error, trace latency at p50 and p99, and trace error rate.
Note the vocabulary collision, because it causes real confusion: Monitoring and Observability are different surfaces here, not synonyms. Monitoring is aggregate charts; Observability is individual records.
Outcomes
What the agent produced that the business values, expressed in money rather than tokens. This is the return side of the cost figure on Overview.
Evaluation
Titled Online Evaluation — scoring of live traffic, as distinct from offline experiments run against a fixed dataset.
Tiles: OVERALL QUALITY as a blended score, COVERAGE as records scored out of records sampled, QUALITY TREND, and EVAL SPEND. Evaluator columns: EVALUATOR, TYPE, SCOPE, SAMPLING, RESULT, COVERAGE, LAST EVALUATED, COST, ADDED BY. Add one with + Add evaluator.
EVAL SPEND is worth knowing about before it surprises you — evaluation is itself a cost, and it is the component of governance spend attributable to this agent. See Where Spend Comes From.
Incidents and Policies
Incidents is this agent's slice of the global incident list, and the fastest path from a violation to the trace that produced it. Policies is a read-only list of the policies applying to the agent, with an Incidents column; click a row to open that policy's Controls tab. Read an empty list as "nothing is checking this agent".
Important considerations
Setup stands in for Overview while pending — An agent always has seven tabs. While it is
Setup Pending,SetupreplacesOverviewand the other six tabs are disabled; direct links redirect toSetup.Observability is agent-scoped — There is no cross-agent trace explorer. Investigations start from an agent.
Tiles follow the period picker — Two numbers from two windows are not comparable. Check the picker before quoting one.
Monitoring ≠ Observability — Aggregate versus individual. The nav does not make this obvious.
An empty tab is usually an empty agent — No traces means no charts, no sessions, no evaluation coverage. Check
LAST TRACEbefore treating a blank tab as a fault.