Automations

The task types

Every task an automation can run, what it needs, and which are Beta.

Last updated August 10, 2026
Reading time 4 min read

A task is one unit of work inside an automation. When you add one, the picker groups the available types into four categories. This article is the reference: what each task does, what you have to give it, and which ones still carry a Beta badge.

Thirteen task types are active today.

Databasin Tasks

Work that runs on Databasin infrastructure.

Task What it does You provide
Pipeline Run an existing Databasin pipeline. The pipeline.
Automation (Beta) Run another automation and wait for it to finish. The child automation.
File Drop Extract data from a file source and load it into a destination. Source + target connections, a query map.
Unzip File Extract a zip archive from a file source into a target file location. Source + target connections, a pattern.
SQL Script Execute SQL against a target connection. Target connection, the statements.
Databasin Notebook (Beta) Execute a saved notebook. The notebook, and optionally a version.
Semantic Model (Beta) Build and deploy semantic models to your warehouse. Target connection, the model definition.

Running one automation from another

The Automation task lets an automation call another one and wait for it to complete — so a long workflow can be composed from smaller, reusable ones instead of one giant stage board.

The picker only offers automations you can actually run: it hides the current automation, inactive ones, and other people's private automations. Self-references and longer cycles are rejected when you save.

Two of the same child in one stage

Putting two tasks that run the same child automation in a single stage doesn't get you two parallel runs — the builder warns you, the same way it does for duplicate pipeline tasks. Split them across stages if you genuinely need both.

Transform Tasks

Task What it does You provide
dbt (Beta) Build, test, or run a dbt project against a target warehouse. Target connection, project source, command, vars.

The default command is build. You can narrow the run with a --select expression and pass variables.

dbt on Databricks authenticates with a token

A dbt task against Databricks needs a Personal Access Token — the Entra ID service-principal path isn't available to it. Databasin warns you when a dbt task is pointed at a service-principal connection.

Intelligence Tasks

Predict, analyze, and report on your data. All three are Beta.

Task What it does You provide
LLM Agent Run a headless agent with a saved skill and produce an artifact. LLM connector, data connection, skill, inputs.
Forecast Forecast a time series in-warehouse and write the result to a table. Target connection, the series and horizon.
OCR Read PDFs with a vision model and extract structured records into a table. Document, LLM, and lakehouse connections + a shape.

LLM Agent

The headless cousin of Databasin One. You pick a skill (the saved instructions for the job), an LLM connector that pins the model, and a data connection the agent's read-only tools query. It runs with its own turn and query caps — defaults are 8 turns and 3 queries — and delivers the artifact through the task's delivery settings.

Forecast

Despite sitting under Intelligence, Forecast spends no tokens. It generates engine-aware SQL — a linear trend with optional seasonality — and runs it through the same SQL executor as a SQL Script task, writing the forecast to a table. It's deterministic and in-warehouse.

OCR

Vision-model extraction from PDFs into a lakehouse table. It's involved enough to have its own walkthrough — see Extract data from PDFs.

Databricks Tasks

Work that runs on your Databricks workspace.

Task What it does You provide
Databricks Job Trigger an existing Databricks job. Target connection, the job.
Databricks Notebook Execute a Databricks notebook. Target connection, notebook path, params, runtime, git.

What Beta means here

A Beta badge means the task works and is supported, but it's newer than the rest and rougher at the edges. The Beta tasks today are Automation, Databasin Notebook, Semantic Model, dbt, LLM Agent, Forecast, and OCR.

Tasks that won't save

Several task types refuse to save when they're missing something that would make the run fail anyway. That's deliberate — a task with a dangling connection used to fail at run time, after the schedule fired:

  • File Drop and Unzip File need both a source and a target connection.
  • Automation needs a child automation selected.
  • OCR needs all three connections plus at least one declared field.

The save error names the specific field, so you don't have to hunt for it.

Where to go next