Skip to main content
Every task declares a type that selects the tool it runs. This page covers the core task types; iteration lives on the Loops page.

type: agent

Runs an agent with a prompt.

Agent task consistency runs

You can enable consistency checking on agent tasks by setting consistency_run to a value greater than 1. This runs the agent multiple times with the same prompt and selects the most consistent output:
This is particularly useful for:
  • Critical calculations where consistency is important
  • Data analysis tasks that may have slight variations in output
  • Queries involving complex logic where you want to verify the agent’s reasoning
Consistency prompt configuration:
  • Automation-level: Add consistency_prompt at the root to apply to all agent tasks
  • Task-level: Add consistency_prompt to a specific task to override the Automation prompt
  • Default: If not specified, uses the built-in evaluation prompt optimized for data analysis (view the default prompt source)
Available template variables:
  • {{ submission_1 }} - First output
  • {{ submission_2 }} - Second output
  • {{ task_description }} - The task prompt
Note: Custom prompts completely replace the default. To adapt the default for your domain, copy the source and modify the specific rules you need.

SQL output mode

Set output.mode: sql to use an agent as a SQL generator. The agent stops once it has SQL — skipping the natural-language answer — and the SQL text becomes the task’s result. SQL compiled from the semantic model or matched from a verified query is returned without running it; SQL the agent wrote itself gets a LIMIT 0 check first. A later step reads the SQL as {{ query.text }}:

type: formatter

Formats the provided template using the outputs of other tasks, then passes the rendered template as output.

type: workflow

Runs another Automation as a sub-step. Use this to break up complex Automations into smaller, easily testable chunks. The variables key allows parameterization by overriding the sub-Automation’s Automation-level variables. This is particularly useful when embedding an Automation task into a loop:

type: semantic_query

Runs a query against your semantic model — measures, dimensions, filters, and ordering by name, with the SQL compiled for you. See Using the Semantic Model for every property.
Looker and Omni integrations add looker_query and omni_query task types — see Looker.

type: conditional

Runs a different set of tasks depending on earlier results. Each branch’s if is a Jinja expression (without the surrounding braces) evaluated against the outputs of previous tasks; the first truthy branch runs. An expression that renders empty, false, 0, or none is falsy.

type: http_request

Makes an outbound HTTP request — for example, to post results to another system’s API. url, headers, body, and form values are Jinja templates rendered against earlier task outputs, and {{ secrets.NAME }} resolves any project secret listed in secrets. Requests to private, loopback, and link-local addresses are always blocked.

type: airway

Runs an Airway ELT pipeline (.airway.yml) as a step. The step completes only once the pipeline’s load has committed, so a following execute_sql task reads fully loaded tables.

Common task options

export: and cache: (which wrote a task’s result to a file in the workspace directory) and agent_ref: __builder__ steps are no longer supported. A .automation.yml that still uses one fails to compile, and the failure names the step. Remove the key or the step; to keep a result, write it to the warehouse with an execute_sql or airway step.
The visualize task type has been retired. Automations that still use it fail with an “unknown task type” error — use an agent task instead.

Next steps

Loops

Iterate over values with loop_sequential

Variables

Parameterize tasks with typed inputs