The agentic data stack

The “modern” data stack was built to serve traditional BI. A new stack is needed for agent-driven analytics.

Designed for agent users

Agents struggle with fragmented context and UI-centric workflows. Graphene consolidates the entire stack into a single repo with everything as code: models, metrics, business context, queries—even dashboards.

See example project →

delay_factors.md

---
title: What makes your flight late?
layout: notebook
---

Every frequent flyer has a pet theory. Pick the right airline. Avoid Chicago. Never fly a 20-year-old plane. These intuitions have just enough truth in them to survive, but they obscure which factors actually move the needle — and by how much.

This is an analysis of FAA data covering roughly 345,000 U.S. commercial flights from 2000 to 2005. For each candidate factor — airline, origin airport, day of the week, and time of day — I measured how much of the variance in individual departure delays each one explains (η², or eta-squared).

## One factor dwarfs the rest

```gsql factor_importance
with
info as (
  from flights where cancelled = 'N'
  select avg(dep_delay) as grand_mean, var_pop(dep_delay) * count() as ss_total
),
hour_groups as (
  from flights where cancelled = 'N' and extract(hour from dep_time)::integer between 5 and 23
  select extract(hour from dep_time) as grp, count() as n, avg(dep_delay) as grp_mean
),
...

The semantic layer, reimagined

Introducing GSQL: a data modeling and query language designed for agents. Agents perform better with GSQL than both regular SQL and traditional BI APIs.

GSQL Model

table orders (
  id BIGINT -- Primary key
  user_id BIGINT
  created_at DATETIME
  status STRING -- One of 'Processing', 'Shipped', 'Complete', 'Returned'
  amount FLOAT -- Amount paid by customer
  cost FLOAT -- Cost of materials

  join one users on user_id = users.id

  revenue: sum(case when status <> 'Returned' then amount else 0 end)
  cogs: sum(case when status <> 'Returned' then cost else 0 end)
  profit: revenue - cogs
  profit_margin: profit / revenue
)

table users (
  id BIGINT -- Primary key
  email VARCHAR
  country_code VARCHAR
)

GSQL Query

select users.country_code,  -- Join is implied via the model
  profit_margin,            -- Invoke modeled measures
  avg(amount) as aov
from orders

Compiled SQL

select users.country_code,
  (
    sum(case when status <> 'Returned' then amount else 0 end)
    - sum(case when status <> 'Returned' then cost else 0 end)
  )
  / sum(case when status <> 'Returned' then amount else 0 end)
    as profit_margin,
  avg(amount) as aov
from orders
left join users on orders.user_id = users.id
group by 1

GSQL Model

table orders (
  id BIGINT -- Primary key
  user_id BIGINT
  created_at DATETIME
  status STRING -- One of 'Processing', 'Shipped', 'Complete', 'Returned'
  amount FLOAT -- Amount paid by customer
  cost FLOAT -- Cost of materials

  join one users on user_id = users.id

  revenue: sum(case when status <> 'Returned' then amount else 0 end)
  cogs: sum(case when status <> 'Returned' then cost else 0 end)
  profit: revenue - cogs
  profit_margin: profit / revenue
)

table users (
  id BIGINT -- Primary key
  email VARCHAR
  country_code VARCHAR
)

GSQL Query

select users.country_code,  -- Join is implied via the model
  profit_margin,            -- Invoke modeled measures
  avg(amount) as aov
from orders

Compiled SQL

select users.country_code,
  (
    sum(case when status <> 'Returned' then amount else 0 end)
    - sum(case when status <> 'Returned' then cost else 0 end)
  )
  / sum(case when status <> 'Returned' then amount else 0 end)
    as profit_margin,
  avg(amount) as aov
from orders
left join users on orders.user_id = users.id
group by 1

Context beyond metrics

Agents need more than metric definitions. They also need a well-organized knowledge base that documents your business. Graphene follows the open standard skill format.

cohort-analysis/SKILL.md

---
name: cohort-analysis
description: Performs cohort analysis for customer behavior patterns over time. Use when analyzing retention, LTV, or other metrics by customer cohorts, or when the user asks about cohorts, retention curves, or customer lifecycle patterns.
---

# Cohort Analysis

Users expect cohort analysis to look a certain way. This guide walks through it.

## Query Pattern

Always group by two dimensions:
1. **Cohort period**, e.g., the month a user's subscription starts
2. **Elapsed time units**, e.g., the number of months that have elapsed since the subscription started

Then aggregate the target metric(s). Filter to the last 12 months if not specified otherwise.

## Visualization Options

### Option 1: Multi-Series Line Chart
...

about-sales-team/SKILL.md

---
name: about-sales-team
description: How to perform analytics for sales and GTM-related areas
---

# Sales

## Overview

The sales team is broken down into sub-teams for inbound sales, outbound sales, and enterprise sales. The Enterprise Sales owns all accounts at the enterprise level (>1,000 employees) and includes both the inbound and outbound directions for this level.

We use Salesforce to manage our sales operations. The most important tables are `dim_salesforce_opportunities` and `dim_salesforce_accounts`.

## Key Concepts

### Opportunity Types
* **New Business:** First-time purchase by a new account.
* **Expansion:** Upsell or cross-sell to an existing account.
* **Renewal:** Continuation of an existing contract.

### Business Sizes
* **SMB:** <100 employees.
* **Mid-Market:** 101 - 1,000 employees.
* **Enterprise:** >1,000 employees.

## Glossary of terms
...

Try Graphene

Run the demo

Clone our example project and ask your coding agent a question about flight data.

View on GitHub

Try it on your data

Let us set up Graphene on your data. Your end users get analytics agents deployed in Slack and in our web app.

Book a call