Skip to main content
Preview Metric Studio holds the context an analytics agent needs to get a business number right: your team’s metrics and saved queries. Agents read them through MCP Gateway before they write SQL, so “revenue last quarter” comes out the same way whoever asks and whichever agent answers.

Why agents need it

Ask an agent “What was revenue last quarter?” without that context, and it writes its own SQL. It picks a table that looks right, guesses the joins, and may leave out the filter that drops cancelled orders. Another agent, or the same one tomorrow, makes different choices and returns a different number. The knowledge that fixes this already exists, in your analysts’ SQL, your BI models, and people’s heads. Metric Studio puts it where agents can find it:
  • A metric is the agreed calculation of one number: its formula, source table, joins, the dimensions it can be split by, its grain, and the filters that must always apply.
  • A saved query is trusted SQL tied to the question it answers. It carries what a definition can’t spell out, such as which table to trust, which edge cases to leave out, and how your business phrases a question.
Agents use both. A metric gives the agreed calculation of a number. A saved query is a worked answer to a similar question that the agent can adapt. On Atlan’s internal talk-to-data benchmark, adding metrics and saved queries took one model from 31.8% to 83.2% accuracy on the same questions. Atlan’s open-source analytics agents playbook describes the method behind that result: build a golden set of real questions first, make the metric the unit of context, and improve from traces.

How it works

Every analytics agent that gives reliable answers needs the same three things:
  • Inputs: the SQL your team already runs, your table schemas, and your BI models.
  • Context: metrics and saved queries, made from those inputs, that the agent retrieves before it writes SQL.
  • An improvement loop: questions with known answers, a score, the traces behind each answer, and versioned changes.
Metric Studio is where the context lives. Atlan Trident supplies the loop. Together they follow this cycle:
  1. Build. Start from the SQL your team already runs in Snowflake or Databricks, Power BI or Tableau model files, or a business question. Your own agent, such as Claude Code, Codex, Cursor, or Atlan Trident Desktop, proposes metric definitions and saved queries, and you approve them before they’re saved.
  2. Review. Each metric explains itself in plain words: what it measures, how it’s calculated, and how it can be split. Studio reads the SQL for problems such as a duplicate calculation and flags them on the metric.
  3. Release. Agents get only released versions. Saving a metric in Studio never changes what agents get until you release it.
  4. Answer. Agents search a workspace for metrics and saved queries, then read the released version over MCP Gateway with the caller’s own permissions. They run the SQL themselves, through their own warehouse access. This works from coding clients and from agents in your data platform, such as Snowflake Cortex Agents and Databricks.
  5. Measure and improve. An eval scores an analytics agent on real questions with known answers, and traces show what it read on each one. A fix becomes a new revision, and the next eval shows whether it helped.

Why it’s part of Atlan Trident

Metrics are one part of a working analytics agent. The other parts, the agent record, its skills, the evals that score it, and the traces of its runs, already live in Atlan Trident. Keeping metrics beside them gives you:
  • One release model. Metrics and saved queries are versioned and released on channels, like skills. You can pin an agent to one version, and withdraw a release that turns out to be wrong.
  • One improvement loop. An eval, the traces behind it, and the metrics the agent read sit in the same workspace, so you can connect a change in score to the definition change that caused it.
  • One access model. Workspace roles decide who can read, edit, and release metrics. Agents reach them through MCP Gateway with the caller’s own permissions.
  • Context that fits together. Metrics govern what is measured and how it is calculated. Ontology supplies shared definitions of your business objects, how they connect, and the policies that apply, which help an agent interpret the results.
Metric Studio and semantic layers. A semantic layer, such as a Snowflake semantic view, a Databricks metric view, or a dbt semantic model, holds the same parts as Metric Studio: table meanings, joins, metrics and dimensions. Metric Studio organizes them around the metric instead of a view, so a metric is defined once, a question can span views, and a fix changes the definition. Keep your semantic layer: Metric Studio can start from the definitions it holds. See Metric Studio and semantic layers.

Get started

Planning a pilot? Your first week puts every step in order, from access to a scored agent, with what you should see after each one.

Build your first metrics

Let your agent propose metrics from query history, review them, and release one.

Set up Analytics in your coding agent

Turn dashboards and query history into metrics, then ask your data questions.

What you can do

Save trusted queries

Keep proven SQL with the question it answers, and link it to metrics.

Write or edit metric

Define a metric by hand, or fix one that gives a wrong answer.

Review agents' suggestions

Apply or dismiss the changes and new metrics that agents propose for your metrics.

Import existing definitions

Already have metric definitions, saved SQL, or a dbt project? Bring them in.

Release to agents

Check a revision, see what changes, and release it.

Connect agents

Agents search metrics and read the released version over MCP.

Retire metric

Withdraw a release, name a replacement, or archive.

Who uses it

Example metric

Every page in these docs uses the same online store, with an orders table. This is what one metric holds: A saved query next to it might answer “What was average order value by region last quarter?” with the exact SQL an analyst trusts.

Key concepts

  • Metric: the agreed calculation of one business number. See How metrics work, and Metric Studio and semantic layers for how it compares with a semantic view.
  • Saved query: trusted SQL and the question it answers. It can link to the metrics it uses.
  • Revision: what every save creates.
  • Release: a revision distributed on a channel, latest by default, with a version number such as 0.1.0. Agents read releases.
  • Release policy: a workspace setting that decides whether a save releases straight away. Studio’s own metric saves never do.

Availability and limits

  • Metric Studio is in preview. An administrator must install the Metric Studio app for your organization first. See Install and manage apps.
  • Anyone with access to a workspace can read its metrics and saved queries. Creating and editing need the workspace Builder or Admin role. Releasing and withdrawing need Admin, or Builder on an item you own.
  • Metric Studio never connects to a warehouse or runs SQL. It reads the stored SQL for a few checks, but releasing doesn’t certify that a number is right.
  • A metric stays in the workspace you create it in. It can’t be moved later.

Reference

Troubleshooting

See Metric Studio errors.