The four kinds
Each agent has a kind that preselects sensible source defaults and an instruction template:How a run works
1
The agent runs
Triggered manually, on record creation, or on record update.
2
It reads its sources
The agent gathers the data you’ve allowed. The record’s own fields, and the web when enabled.
3
The AI produces a value
The model returns a candidate value for the property.
4
The value is checked and written
Custral verifies the value fits the property’s type. If it fits, it’s written to the cell. If it doesn’t, the run is marked failed with a clear message. The unvalidated text is never written as a fallback.
Triggers vs manual run
An agent runs manually when you click Run, or automatically on create and on update. The full trigger reference is on Triggers.Output validation
Every value the AI produces is checked against the target property’s type before it’s written. A select option must exist, a date must parse, a number must be numeric. On failure, the run is marked failed with a descriptive error, and the unvalidated text is never written as a fallback. See Troubleshooting for how to read these errors.Supported property types
The image type is special: the AI never returns the picture directly. Instead it ranks real candidate images from a web image search, and Custral validates, downloads, and attaches the chosen image to the record.