Agent Observability overview

Kubit Agent Observability is Agent Observability with Behavior Analytics. It reads the OpenTelemetry traces your agents already produce, labels every span with Polaris and groups every agent run by its execution path, so you can see which runs went wrong and how often each path happens.

What Agent Observability does with your traces

Agent Observability adds three things on top of raw spans.

Feature	What it does	Example question it answers
Polaris	Kubit’s in-house small language model labels every span with what it did and whether it worked. Each span gets labels in under a second, with no sampling.	Which tool calls returned 200 but picked the wrong parameter?
Flows	Groups every agent run by execution path (which tools ran, in what order) and draws all paths in one Sankey.	What share of refund conversations loop on `search_orders` three or more times?
Triggers	Compares a metric across model versions or user segments and tells you when it changes.	Did the Misstep rate on `book_appointment` go up after the model upgrade?

Polaris annotations come in four types: Action, Exception, Misstep and Warning. Concepts defines each one.

How traces get into Kubit

Kubit reads agent traces from one of three places.

Where your traces are	What you do	Start here
In your app, not stored anywhere yet	Install the Kubit OpenTelemetry SDK (`kubit-otel` for Python, `@kubit-ai/otel` for Node) and send spans with an API key.	Quickstart
Already in Langfuse or Braintrust	Add `KubitSpanProcessor` next to your existing processor. Your current tool keeps working.	Set up with your coding agent
Already in your warehouse	Connect Databricks, Snowflake, BigQuery or ClickHouse. Kubit queries the trace tables where they are.	Bring your own warehouse

No data silo to secure. Traces retained in your control.

Who uses Agent Observability

Reader	Typical use
AI product engineer	Instruments the agent, finds the execution paths that fail and checks a fix against the next week of runs.
Product manager or analyst	Reads Flows to see which agent behavior leads to a completed task, a support ticket or a cancellation.
Coding agent	Wires Kubit into a repo with `/kubit-integrate`, then queries traces through the Kubit MCP server.

Quickstart

First traces from a Python agent in about 10 minutes.

Concepts

Users, sessions, traces, spans, execution paths and Polaris annotations.

Set up with your coding agent

Let Claude Code detect your tracing setup and wire Kubit in.

Is my data flowing?

Check each step when traces don’t show up.

​What Agent Observability does with your traces

​How traces get into Kubit

​Who uses Agent Observability

​Related