JOIN THE WAITLIST
Agent Observability with Behavior Analytics
Kubit’s in-house small language model labels what every span did and whether it worked. Every agent run then visually grouped by execution path across LLM, RAG and tool steps in a Sankey. Filter by user, version or cohort to see which paths fail and what changes trigger them, all in your warehouse.
Upload a trace file to get started
Every span labeled and visually grouped
Traces retained in your control
THESE TEAMS ALREADY RUN KUBIT PRODUCT ANALYTICS IN THEIR WAREHOUSE
PROBLEM
Tens of thousands of runs a week, opened one trace at a time. The failures that matter get found by customers first.
The dashboard is green yet the product is broken
Every span returns OK, so quality failures never reach a screen. The customer finds them first.
No way to tell a big problem from a loud one
Every occurrence arrives as its own trace. The fix list gets ordered by who complained, not by what is costing the most.
Agent behavior and business data live apart
The warehouse holds the traces and the revenue tables. Every question that joins the two becomes a data-engineering ticket.
PLATFORM
Query the agent traces already in your warehouse. Nothing to install.
The traces are already in your warehouse. Kubit queries them there. No SDK, no pipeline.
Databricks Native
Reads MLflow traces and evals in Unity Catalog.
No SDK to import
No pipeline to build
Snowflake Native
Reads AI_OBSERVABILITY_EVENTS and GET_AI_EVALUATION_DATA.
Cortex Agents traces, read in place
Evaluation data lands beside the traces
BigQuery & ClickHouse Native
The same native read, now in BigQuery and ClickHouse.
Queries run in your own account
No second copy to secure
Trace File Upload
Same analysis, no warehouse needed.
Try it before connecting anything
Same labels, same clustering, same analytics
Outcome Joins
Joins agent actions to user activity, a/b tests and campaigns.
See which execution paths actually convert
Tie a pattern to a business outcome, not just an error rate
One query, because it is one warehouse
WHY KUBIT
Label each span, group by execution path, in your warehouse.
Every span is labeled, including OK ones.
Polaris, Kubit’s in-house small language model, labels what every span did and whether it worked. Flows then groups all runs that called the same tools in the same order.
Every failure comes with its triggers.
For each failing execution path, Triggers shows the change deployed before it and the attributes, like model version or user segment, most common in its failed runs.
Directly queries your own warehouse.
No SDK, no pipeline. Your warehouse already holds user activity, A/B tests and campaigns, so you can tie what the agent did to the outcomes that followed.
Your traces stay in your warehouse.
Kubit queries your agent traces in place, inside your own Databricks, Snowflake, BigQuery or ClickHouse account. There's no data silo to secure.
Your Account, Your Data · Bring Your Own Warehouse
QUERY TRACES IN DATABRICKS, SNOWFLAKE, BIGQUERY AND CLICKHOUSE
FAQ
Questions We Often Get
We already have an observability tool. What is Kubit for?
Kubit works on the traces already stored in your warehouse, and standard OTel collection can write them there directly. Try this on your current tool: ask for last week's ten most common execution paths, ranked by run count, with the change and attributes behind each one. Then see what comes back.
Why not just use what is built into Databricks or Snowflake?
If reading one trace at a time is enough, use them. Kubit queries the same tables, so nothing moves. The difference shows when you ask what's happening across all your runs, with funnels, cohorts and trigger signals. For example, MLflow's issue detection only clusters traces that failed a check, so your successful runs never make it into the picture. Kubit labels every span and clusters every run, passing and failing, by the shape of its path.
Could we not build this ourselves?
Collecting traces takes about an hour, so that part is easy. Three pieces are harder to build yourself: one diagram of every run, clustered by the shape of its path; a fine-tuned model that labels every span; and behavior analytics with funnels, cohorts and trigger signals. Then all three have to keep running six weeks later, after the person who built them has moved on.
What does labeling every span cost us?
Nothing extra. Polaris Evaluator is a small language model we built, not an LLM-as-a-judge, and it's included in every plan. You pay per trace, not per span, so labeling every step of a run doesn't add to the bill, and there are no LLM credits to buy.
Should we trust a cluster label enough to act on it?
The combination, in one place. Kubit labels every span, clusters every run by the shape of its execution path and draws them as one Sankey, finds the signals behind each pattern, and ties it all to your business data, in your own warehouse. Try the demo above, then join the waitlist to run it on your own traces.
It is a crowded space. What is actually different?
One mechanism, not a superlative. Kubit groups runs by the shape of the execution path, draws every shape in the population as one diagram, hands over the signal behind a pattern, and joins it to what the business got, in the same warehouse. This is early access and the agent product is new. Open the demo and judge the mechanism.
Do I have to install another SDK?
No. Start by uploading a trace file. In production, Kubit works on agent traces already in your cloud data warehouse, collected with standard OTel, Databricks MLflow Tracing or Snowflake Cortex Agents.
Coming Soon in the Roadmap
The first cohort is closed. Join the waitlist for the next one.
THESE TEAMS ALREADY RUN KUBIT PRODUCT ANALYTICS IN THEIR WAREHOUSE