Agent Observability Terms
| Term | Definition |
|---|---|
| Agent Observability with Behavior Analytics | What Kubit is: Agent Observability features for agent traces, on top of Product Analytics reports that run on agent traces and clickstream events, separately or joined |
| Agent Observability | The set of features built for agent traces: Polaris, Flows and Triggers. See Agent Observability overview. |
| Execution path | The ordered sequence of spans in one trace: which tools and model calls ran, in what order. Two traces with the same steps share one execution path, even if their inputs differ. See Concepts. |
| Polaris | Kubit’s in-house small language model. It labels and scores every span, in under a second per span, with no additional LLM cost or sampling. See Polaris. |
| Polaris annotations | The labels and scores Polaris adds to every span |
| Label | Classifier output, such as wrong_tool or call_loop |
| Score | Eval output: a number, a true or false, or a value from a list you define |
| Evaluator | The page in the app where Polaris answers system questions and your custom questions about each conversation |
| Action | Annotation type: the span did its job, such as an LLM call, a retrieval or a tool call |
| Exception | Annotation type: the call failed, with an exception, a 4xx or 5xx response, a timeout or no network |
| Misstep | Annotation type: the call ran, often with an OK status, but did the wrong thing, such as wrong_tool or malformed_args |
| Warning | Annotation type: the span is a risk even if it worked, such as a loop, high cost, high latency, high frequency, detected PII or a spike |
| Behavior | Annotation type: what the agent and the user did and said, such as intent, sentiment or hallucination |
| Flows | The view that groups every agent run by execution path and draws all paths in one Sankey with a count per path. See Flows. |
| Triggers | The feature that compares a metric across model versions or user segments and shows when it changes. See Triggers. |
Data Terms
| Term | Definition |
|---|---|
| User | Subject: the end user who started the trace, identified by User ID |
| Session | Subject: a group of related traces, such as one conversation, identified by Session ID |
| Trace | Subject: one end-to-end agent run or request, identified by Trace ID |
| Span | Subject: one timed operation inside a trace, such as an LLM call, a tool call or a retrieval, identified by Span ID |
| Clickstream event | An event your web or mobile app sends about user activity, collected by a customer data platform such as RudderStack or Segment |
| Workspace | One set of traces or events, their Polaris annotations and their reports. An organization can have several. |
| Origin | Where a workspace’s data comes from, picked when you create it: Sandbox, MLflow Tracing, Cortex Agents, Google ADK, Langfuse, Warehouse, Agent or User. See Set up your organization. |
| Sandbox | An origin for uploading trace files for a quick test, with no pipeline |
| Extraction mapping | The rules that decide which span attributes fill in the User ID, Session ID and other fields of an Agent workspace |
| Modeler | Where you map warehouse tables to fact tables, subjects and fields. See Modeler. |
Product Analytics and Tools
| Term | Definition |
|---|---|
| Product Analytics | The foundation of Kubit: reports, cohorts, dashboards and Lumos on any behavior events, agent traces and clickstream events alike |
| Path report | The Product Analytics report of event sequences. It isn’t Flows, and is never called Flow. See Path report. |
| Lumos | Kubit’s in-app AI agent: ask a question in plain language and get a report back. See Lumos. |
| Kubit MCP server | Lets Claude Code, Cursor, Codex and ChatGPT build Kubit reports and inspect traces from a chat. See Kubit MCP Server. |
| Agent plugin | The Claude Code and Cursor plugin whose /kubit-integrate skill wires Kubit into a repo. See Agent plugin. |