Automations

An automation combines instructions, a model, visibility, mutation policy, attachments, and triggers. Start with a small task in a development project before enabling unattended execution.

Create your first automation

1. Open AI → Automations → New Automation.
2. Give it a clear name, such as Review catalog quality.
3. Choose the appropriate visibility: personal, project, or a permitted space.
4. Write bounded instructions: “Inspect the products, identify missing category links, and summarize them. Do not change records.”
5. Select a model and retain Mutations: Require approval for the first exercise.
6. Attach any required skills, custom tools or plugin tools. Only attached tools are callable; discovering the catalog is not enough.
7. Save the configuration. New automations start Inactive; enable configured triggers deliberately after testing.
8. Use the manual run action, then inspect Run History and the summary.

Manual execution is separate from the Add Trigger menu. The configured trigger choices are Scheduled, Webhook Triggered, and Entity Events.

Plugin tool calls follow the same mutation policy: with Require approval, calls that are not clearly low-risk wait in Approvals; with Auto-approve, or for a tool a reviewer chose to always allow, they run without review. Custom tools can have external effects outside this approval flow. Review their authorization and behavior separately.

Add capabilities incrementally

Attach skills for repeatable instructions, plugin tools for connected services such as Gmail or Shopify, and custom tools for server-side code. Send to Slack and Send to Microsoft Teams post results to channels. Attached skills are loaded; other visible skills can be discovered. Tool availability is narrower: the automation must have the custom tool or plugin tool attached.

After a read-only run works, test a small proposed change and its approval flow. Then configure triggers. Check credits and loop protection before enabling entity-event work.

Use list filters for status, visibility, and credit context when locating an automation. A successful execution badge must be read together with its task summary.

If a run reports that the API cannot aggregate or rank the data the instructions require, treat that as a schema iteration rather than an automation failure.

For the complete exercise, follow Create an automation.