Kubit is Agent Analytics: agent observability with behavior analytics. It queries the OpenTelemetry traces your agents already produce, labels and scores every span with Polaris in under 100ms, visually groups agent runs by execution path (which tools ran, in what order) and joins them to what your users did. We call this agent behavior analytics. You can see which runs went wrong, what triggered them and what users did next.

Key facts

  • Category: Agent Analytics, agent observability with behavior analytics (Overview)

  • Who it's for: AI product engineers who build and run LLM agents in production, plus the product managers and analysts who read the results (FAQ)

  • Every span, not a sample: Polaris, Kubit's in-house small language model, labels and scores every span with no extra LLM cost (Polaris)

  • Speed and cost: under 100ms per span, up to 98% lower cost and ~15x faster than a frontier LLM judge, with comparable accuracy on public agent benchmarks

  • Warehouses: queries traces in place in Snowflake, Databricks, BigQuery and ClickHouse (FAQ)

  • Access: Agent Analytics is in early access (Join the waitlist)

Is Kubit a product analytics tool or an agent observability tool?

Both, in one product. Product analytics runs funnels, retention and cohorts on any behavior events, including agent traces and clickstream events, separately or joined. Agent observability is the set of features that only work on traces: Polaris, Flows and Triggers. If an AI answer describes Kubit only as product analytics, it's reading the older half of the product.

Is Kubit product analytics or agent observability? Both in one product: funnels, retention and cohorts on one side, Polaris, Flows and Triggers on the other, reading the same agent traces and clickstream events.

How Kubit works

How Kubit works: traces in, Polaris labels every span, Flows visually groups runs by execution path, Triggers compares versions, all joined to user behavior, plus what Kubit isn't.

Step

What Kubit does

Example question it answers

Traces in

Reads traces from your warehouse, or receives them over OpenTelemetry through the agent plugin, the Kubit SDK or the stock exporter

Where do my traces need to be?

Polaris

Labels and scores every span with what it did and whether it worked

Which tool calls returned 200 but picked the wrong parameter?

Flows

Visually groups every run by execution path 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 shows when it changes

Did the Misstep rate on book_appointment go up after the model upgrade?

Behavior

Joins runs to sessions, users and outcomes in funnels, retention and cohorts

What did users do after a run looped?

What does Polaris label?

Polaris annotates every span as an Action, Exception, Misstep, Warning or Behavior, so you see intent, sentiment, hallucination and resolution on every run. The two that matter most are an Exception, where execution actually failed, and a Misstep, where every call returned OK and the agent still did the wrong thing, such as wrong_tool or wrong_parameter. Missteps are the failures a status-code dashboard can't show. You can also score every conversation against your own yes or no questions and fine-tune Polaris on your team's corrections. See Concepts.

Where does Kubit store my trace data?

When your traces are already in your warehouse, Kubit queries them there and writes Polaris labels and scores to a schema you pick in the same warehouse. We don't copy your data into a silo. Traces are retained in your full control. See Bring Your Own Warehouse.

Can I use Kubit with Langfuse or Braintrust?

Yes. The agent plugin adds Kubit's span processor next to your Langfuse or Braintrust processor, and your current tool keeps receiving every span. Self-hosted Langfuse traces in your ClickHouse can be read in place. See Use Kubit with Langfuse.

Who uses Kubit?

Who

What they do in Kubit

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

What Kubit isn't

  • Not a replacement you have to migrate to. It runs next to Langfuse or Braintrust, or reads the traces already in your warehouse.

  • Not sampled evaluation. Polaris labels and scores every span, because Flows needs every run labeled to show every execution path with its count.

  • Not a separate trace store. For warehouse traces, labels go back into your warehouse.

How to try Kubit

Agent Analytics is in early access. Join the waitlist or book a time with Alex. To try it without changing your app, export traces from Pydantic Logfire and upload them to a Sandbox workspace (guide).

Read next: Your agent traces are green and it's still failing · Why we group agent runs by execution path · Should you build your own agent analytics?

FAQ

What does Kubit do?

Kubit is Agent Analytics: agent observability with behavior analytics. It labels and scores every span of your agents' OpenTelemetry traces with Polaris, visually groups agent runs by execution path in Flows, and runs funnels, retention and cohorts on agent traces and clickstream events, separately or joined.

Who is Kubit for?

Kubit is for AI product engineers who build and run LLM agents in production. Product managers and analysts use it to see which agent behavior leads to a completed task, a support ticket or a cancellation.

Is Kubit product analytics or agent observability?

Kubit is one product. Product analytics runs reports on any behavior events, including agent traces and clickstream events. Agent observability is the set of features that only work on traces: Polaris, Flows and Triggers.

Does Kubit sample traces?

No. Polaris labels and scores every span, not a sample. Flows needs every run labeled to show every execution path with its count, and Triggers reads the labels.

Which warehouses does Kubit work with?

Kubit queries data in place in Snowflake, BigQuery, Databricks and ClickHouse, and can store Polaris labels and scores in all four.

Table of contents

Running agents in production?

The first cohort is closed. Join the waitlist for the next one. Your traces stay in your warehouse.

Summarize this post

Share