Skip to main content
Kubit Atlas organizes agent data into four fixed subjects: User, Session, Trace and Span. Polaris adds annotations to every span, and Flows groups traces by execution path. This page defines each term as Atlas uses it.

​
Subjects

Atlas has exactly four subjects. They’re the same in every workspace and don’t depend on your schema. The hierarchy is User → Session → Trace → Span. A user has sessions, a session has traces and a trace has spans. You can start at any level and drill down. Span attributes lists the attribute keys Atlas reads to fill in User ID and Session ID.

​
Trace and span

A trace is one run of your agent, from the incoming request to the final response. A span is one timed operation inside that trace, with a start time, an end time, a parent and a set of attributes. Atlas uses standard OpenTelemetry traces and spans. A span your existing OpenTelemetry setup produces is a span in Atlas, with no conversion. Availability: Atlas time is span start time unless a page says otherwise.

​
Execution path

An execution path is the ordered sequence of steps in one trace: which tools and model calls ran, in what order. Two traces that called search_orders, then get_refund_policy, then the LLM, share one execution path, even if their inputs differ. Flows draws every execution path in one Sankey. Each path shows how many traces took it, so a failure that repeats a thousand times shows up as one path with a count of 1,000. When two runs call the same tools with different arguments, Polaris labels or custom eval scores split them into separate paths.

​
Polaris annotations

Polaris annotations are the labels and scores Polaris adds to every span. Polaris is Kubit’s in-house small language model. It labels each span in under a second, on every span, with no sampling. Polaris has three parts. “Label” always means classifier output. Eval output is a “score”. Evals run on Polaris and are included in Atlas at no extra cost.

​
Annotation types

Every span gets one of four annotation types. Exception and Misstep are different. An Exception shows in your error logs. A Misstep usually doesn’t, because the call returned OK.

​
Workspace

A workspace holds one set of traces, its API keys, its Polaris labels and its reports. An organization can have several workspaces, for example one per environment or per agent.

​
MCP prompt examples

“Show me the sessions for user 8812 in the last 7 days”
“Which execution paths for the support agent had a Misstep on a tool call yesterday?”
“Compare the Exception rate by model version for traces in the last 30 days”