Unify Logo Footer.svg
Platform Tools
Logo
Observability - global traces

Observability - global traces

Logo

3 mins READ

The platform-level trace browser: every AI execution trace in one table, with a step-by-step timeline for each.

Overview

Inside a project there's an Observability screen with a single Traces tab. Unlike an agent's own Observability tab — which shows sessions and traces for that one agent — this is the global trace browser: it lists execution traces across agents in one place, including traces reported into the platform by external agents. Two things to know: it has no side-navigation entry today, so you reach it by URL (the project's /observability/traces path or a link from another screen), and the list is time-scoped rather than filtered to the project you opened it from.

Where: Project › Observability › Traces

Screenshot TODO — Observability — global traces

Traces table

A table of traces with Trace IDInputOutputLatency, and Start Time columns. A Date Filter (defaulting to the last 30 days, with presets) scopes the list — the table only loads once a date range is set — and more rows load as you scroll (10 at a time). Attribute filters are available on top of the date range. Click a row to open the trace's timeline in a drawer.

Where: Observability › Traces

Seeing when I open a trace

A wide, resizable drawer titled with the trace's name and ID. On the left is the trace timeline — the ordered, nested steps the execution ran; on the right, the detail panel for whichever step is selected, with its Input and Output payloads and a strip of facts about the step: its status, the model used, the tool invoked, latency, cost, tokens (input/output, plus cache reads and writes where applicable), and timing.

Where: Observability › Traces, trace drawer

FAQs

What kinds of steps appear in a timeline?

Each step is typed and color-coded: Query (the incoming request), LLM (a model call), EmbeddingRetrieval and Vector Search (knowledge lookups), RerankerTool (shown with the actual tool's name and icon), Agent ExecutionGuardrail (including Request Guardrail and Response Guardrail), Re-phraser, and PII Mask. Walking the steps top to bottom is how you diagnose a bad answer: see what was retrieved, which tool ran, what each model call cost, and where things went wrong.

Where: Observability › Traces, trace timeline

Can I jump from a trace to an automation run?

Yes — when a step in the timeline was an automation invoked as a tool, its detail panel shows a note — "An automation was executed as a tool." — with a View Run button that opens that automation's run in a new tab, so you can continue debugging in the automation's own run view with full node-by-node detail.

Where: Observability › Traces, step detail › View Run