This post is about agents whose traces land in the warehouse through the platform's own tooling: MLflow Tracing on Databricks, Cortex Agents on Snowflake and Google ADK on BigQuery. Each one writes every trace into the warehouse for you, so you can analyze the traces where they are. All three let you query the traces with SQL and open them in a built-in view or dashboard. Kubit queries the same trace tables in place, labels and scores every span with Polaris, visually groups runs by execution path (which tools ran, in what order) and writes its labels and scores back 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.

Key facts

  • Databricks: MLflow traces are stored in Unity Catalog as Delta tables in OpenTelemetry format, queryable from the MLflow UI, the Python SDK, SQL and AI/BI dashboards (Databricks docs, updated 12 Nov 2026)

  • Snowflake: Cortex Agents monitoring data is stored in SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS and shown in the Observability tab (Snowflake docs)

  • BigQuery (Google ADK): the BigQuery Agent Analytics plugin writes agent events (LLM requests and responses, tool calls, tool errors, agent responses) to an agent_events table, with per-event views and a Looker Studio template or Looker Block on top (ADK docs)

  • Kubit: reads MLflow traces in Databricks, Cortex Agents traces in Snowflake and Google ADK events in BigQuery in place, with nothing to install in your app (MLflow, Cortex Agents, Google ADK preset)

  • Where labels go: four KUBIT_ tables in a schema you pick, next to the traces (Bring Your Own Warehouse)

  • Any other agent framework: send OpenTelemetry traces to Kubit with a short OTel Collection setup, and Kubit writes them to Snowflake, Databricks, BigQuery or ClickHouse (Manual OpenTelemetry, Bring Your Own Warehouse)

What do Databricks, Snowflake and BigQuery give you for agent traces?

All three give you a good place to open one run end to end and query it with SQL.

On Databricks, MLflow Tracing stores traces in Unity Catalog tables in OpenTelemetry format, so each span keeps its attributes and parent-child structure. You open a trace in the MLflow UI, and because the tables are SQL-native Delta tables, you can query them and build AI/BI dashboards on them. You need a Unity Catalog workspace and MLflow 3.14 or later.

On Snowflake, Cortex Agents log their monitoring data to SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS. The Observability tab lists conversation threads with the first input, user feedback and thread length, along with execution traces, and the same data is readable with table functions. Evaluations score an agent on a dataset before or after deployment. For agents built outside Cortex, Snowflake points to TruLens.

On BigQuery, Google ADK agents use the BigQuery Agent Analytics plugin. It writes each agent event, such as an LLM request, a tool call or a tool error, as a row in an agent_events table, with the trace and span IDs, latency and error message. The plugin creates a view per event type, and Google provides query recipes, a Looker Studio template and a Looker Block for dashboards.

What does Kubit add on top of the same tables?


Databricks built-in

Snowflake built-in

BigQuery with Google ADK

Kubit, on the same tables

Where traces live

Unity Catalog Delta tables

AI_OBSERVABILITY_EVENTS

agent_events

Stay where they are

Open one run

MLflow UI

Observability tab

Your own query or Looker dashboard

Activity

Query with SQL

Yes

Yes, through table functions

Yes, plus a view per event type

Yes, labels sit in your schema

Label and score every span

Not covered in the tracing docs

Not covered in the monitoring docs

Not covered in the plugin docs

Polaris, every span, no sampling

Visually group runs by execution path

Not covered

Not covered

Not covered

Flows, one Sankey for all runs

Compare across model versions

Not covered

Not covered

Not covered

Triggers

Join to user behavior

Your own SQL

Your own SQL

Your own SQL

Funnels, retention and cohorts

"Not covered" means the pages linked above don't describe it, not that it can't be built.

How Kubit reads traces in your warehouse

The data path from trace tables in Snowflake or Databricks to Flows, with a table of what the built-in views cover and what Kubit adds.
  1. You create an identity for Kubit with SELECT on the tables it should read, and a compute warehouse sized for it on Snowflake and Databricks. On BigQuery it's a read-only service account. Kubit's queries run on compute you control and can monitor.

  2. You create a workspace with the MLflow Tracing origin on Databricks, the Cortex Agents origin on Snowflake or the Google ADK origin on BigQuery, and apply the preset. Nothing changes in your app.

  3. Polaris labels and scores every span. An Exception is a call where execution failed. A Misstep is a call that returned OK while the agent did the wrong thing, such as wrong_tool or wrong_parameter.

  4. Labels go back to your warehouse, in KUBIT_FACT_TRACES, KUBIT_FACT_SPANS, KUBIT_FACT_CONVERSATIONS and KUBIT_FACT_TRACE_OUTCOMES, in a schema you pick.

  5. Flows, Triggers and behavior reports read the labeled runs: every run visually grouped by execution path, metrics compared across model versions, and runs joined to the user tables already in the warehouse.

Does Kubit copy my traces?

Where your agent data lives with Kubit: trace tables stay in your warehouse, Kubit reads them in place and writes Polaris labels and scores to KUBIT_ tables in a schema you pick, in Snowflake, Databricks, BigQuery or ClickHouse.

No. Kubit signs in with the identity you create, reads only the tables you grant and writes only to locations it owns: the Polaris results schema and, on Snowflake, a small KUBIT database. Every table it creates starts with KUBIT_, so it doesn't collide with yours. See Bring Your Own Warehouse.

What if my agent isn't built on MLflow, Cortex Agents or Google ADK?

Any agent can get its traces into your warehouse. Point the stock OpenTelemetry exporter, or the Kubit SDK for Python or Node, at Kubit's OTLP endpoint with a short OTel Collection setup, and Kubit writes the traces to your Snowflake, Databricks, BigQuery or ClickHouse. From there everything in this post applies: Polaris labels and scores every span, and Flows visually groups the runs. Setup: Integrate with manual OpenTelemetry and Bring Your Own Warehouse.

What about ClickHouse and Redshift?

ClickHouse works the same way: Kubit queries traces in place and can store Polaris results there. Redshift isn't supported for agent traces yet.

When are the built-in views enough?

If you mostly open one run at a time to debug it, or watch token and latency charts, the MLflow UI, the Snowflake Observability tab and the ADK Looker dashboards do that job. Kubit is for the next question: how many runs failed while every call returned OK, which execution paths they took, and what users did afterwards.

How to try it

Agent Analytics is in early access. Join the waitlist or book a time with Alex. Setup guides: Use Kubit with MLflow Tracing · Use Kubit with Cortex Agents · Google ADK preset.

Read next: What is Kubit? · Should you build your own agent analytics? · Why we group agent runs by execution path

FAQ

Can I analyze agent traces that are already in Databricks?

Yes. Databricks stores MLflow traces in Unity Catalog Delta tables in OpenTelemetry format, which you can open in the MLflow UI or query with SQL. Kubit reads the same tables in place through an MLflow Tracing workspace, with nothing to install in your app.

Can I analyze Snowflake Cortex Agents traces?

Yes. Cortex Agents store monitoring data in SNOWFLAKE.LOCAL.AI_OBSERVABILITY_EVENTS, shown in the Observability tab. Kubit reads Cortex Agents traces in place through a Cortex Agents workspace and labels and scores every span with Polaris.

Can I analyze Google ADK traces in BigQuery?

Yes. The ADK BigQuery Agent Analytics plugin writes agent events to an agent_events table in BigQuery. Kubit reads that table in place through a Google ADK workspace and labels and scores every span with Polaris.

Does Kubit copy my traces out of the warehouse?

No. Kubit queries your traces directly in your warehouse and writes Polaris labels and scores to KUBIT_ tables in a schema you pick. We don't copy your data into a silo. Traces are retained in your full control.

Which warehouses does Kubit support for agent traces?

Databricks, Snowflake, BigQuery and ClickHouse. Agents built on any framework can send OpenTelemetry traces to Kubit, which writes them to any of the four. Redshift isn't supported for agent traces yet.

Where do Kubit's queries run?

On Snowflake and Databricks, on a compute warehouse you create and size for Kubit, so you control and can monitor the cost.

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