A row of five blue wireframe cube groups: stacked blocks, a tiled grid with marked tiles, linked columns, a cluster with one raised block, and descending columns.
A row of five blue wireframe cube groups: stacked blocks, a tiled grid with marked tiles, linked columns, a cluster with one raised block, and descending columns.

What is Kubit? Agent Analytics, from trace to user

Alex Li

|

Kubit labels and scores every span of your agent traces, visually groups runs by execution path and joins them to what your users did. Here is what it does, who it's for and what it isn't.

A single line of blue wireframe blocks, one of them marked with a dashed ring.
A single line of blue wireframe blocks, one of them marked with a dashed ring.

Misstep vs Exception: why your agent fails when its traces are green

Alex Li

|

An agent can fail while every trace stays green. Here is the difference between an Exception and a Misstep, and how to find Missteps across every run.

Two parallel rows of tiles: a lit front row with dots and dashed rings, and a faint row behind it.
Two parallel rows of tiles: a lit front row with dots and dashed rings, and a faint row behind it.

Langfuse alternatives for agent analytics in 2026

Alex Li

|

Six Langfuse alternatives for 2026, compared on open source, self-hosting and what each one bills for, and where Kubit fits next to Langfuse.

Small clusters of blue wireframe tiles beside a flow of linked columns.
Small clusters of blue wireframe tiles beside a flow of linked columns.

Braintrust alternatives for agent analytics in 2026

Alex Li

|

Six Braintrust alternatives for 2026, compared on open source, self-hosting and pricing, and the difference between grouping runs by summary and by execution path.

A layered stack of blue wireframe blocks with a lit column running through it.
A layered stack of blue wireframe blocks with a lit column running through it.

Agent analytics on traces already in Snowflake, Databricks or BigQuery

Alex Li

|

Databricks, Snowflake and BigQuery already store your agent traces. Here is what their built-in views cover, and what Kubit adds on the same tables.

A wide field of tiles with a small block lit in one corner. A few dark tiles further back carry a faint dashed mark.
A wide field of tiles with a small block lit in one corner. A few dark tiles further back carry a faint dashed mark.

Should you build your own agent analytics?

Alex Li

|

A script over your trace export answers the first questions about your agent. Here is what it takes to keep answering the next ones, and when building your own is the right call.

A row of tiles read left to right as time. One tall lit block marks a change; tiles after it light up, and dashed lines run from them back to the change.
A row of tiles read left to right as time. One tall lit block marks a change; tiles after it light up, and dashed lines run from them back to the change.

How to find what changed right before an agent pattern started

Alex Li

|

A quality drop sends you through prompts, models, config and commits by hand. Reading the signal off the whole pattern, rather than one trace, gets you to the cause faster.

Four wireframe columns standing in a ring on a tiled floor, joined by a dashed loop. The front column is lit.
Four wireframe columns standing in a ring on a tiled floor, joined by a dashed loop. The front column is lit.

Green traces, broken agents: what we showed with Sparta

Alex Li

|

In our live session with Sparta, an agent sent a refund to the wrong customer while every span stayed green. We traced it to one release, reproduced it in Sparta Worlds and turned it into a test that runs on every release.

Why we label every span instead of a sample

Alex Li

|

Sampling the judge is a rational answer to per-call prices. It leaves most runs without a label, and those are the runs you need to group a pattern or find its cause.

Thin lines, one per agent run, cross and sort into three blocks. The top block is lit and has a loop arrow over it.
Thin lines, one per agent run, cross and sort into three blocks. The top block is lit and has a loop arrow over it.

Why we group agent runs by the shape of their execution path

Alex Li

|

The same cause gets investigated by different people on different days. Group every run by its execution path, which tools in what order, and each repeat joins one shape.

Every trace looks green and the agent is still failing

Alex Li

|

Every span returns OK and the user still got the wrong answer. The failures that cost trust don't throw. Give every span a label so they have a name you can count.

Nobody should have to open hundreds of agent traces one at a time

Alex Li

|

At a hundred runs, opening every trace is diligence. Somewhere before a thousand it turns into a sample nobody chose. One picture of all agent runs fixes that.