Field Agents
Generate typed text, selections, numbers, images, videos, and linked records directly inside collection fields.
Field Agents add AI generation to an ordinary typed collection field. The underlying field keeps its normal sorting, filtering, forms, exports, and workflow behavior — only how its value gets filled in changes.
Choose Field agent from the field picker, or open a compatible standard field and select Convert this field. Converting back to a standard field keeps the current values and immutable run history.
Output kinds
Text, selection, and number outputs
Long text, single select, multiple select, number, percent, and currency fields generate a plain value from a prompt — no configuration beyond the prompt and the fields it references.
Example: a Summary (Long text) agent referencing @Description and
@Reviews writes a two-sentence pitch each time either source changes.
Product images
Image agents either generate a new image from scratch or edit an existing one, referencing up to 8 other fields as visual inputs (editing requires at least one reference). Output size is either a fixed aspect ratio (1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16) or explicit pixel dimensions.
Example: a Product shot (Image, generate, 1:1) agent referencing
@Product name and @Description produces a studio-style product photo; an
Edited shot (Image, edit) agent referencing that shot plus a
@Brand backdrop file field swaps in the brand background.
Product videos
Video agents pick one of five operations, each expecting different source fields:
| Operation | Needs |
|---|---|
| Text to video | Just the prompt |
| Image to video | One image field to animate |
| First / last frame | Two image fields — start and end frames |
| Extend | One existing video field to continue |
| Remix | One existing video field to restyle |
Every video agent also sets duration (1–30 seconds), resolution (720p or
1080p), aspect ratio, whether to generate audio, and an optional
provider-specific quality tier (for example a model's std/pro mode).
Pricing is metered by the actual seconds, resolution, and audio setting used.
Example: a Product video (Video, image-to-video, 1080p, 8s) agent
referencing the generated @Product shot produces a short orbiting
camera-motion clip.
Deep Match relations
Deep Match writes real linked-record IDs instead of a computed lookup: it searches a target collection's fields for semantic matches to the agent's instructions and links the results. Choose single (writes one match) or multiple (up to 20). A minimum match score and an optional reranking pass control precision.
Example: a Best supplier (Deep Match, single, target: Suppliers) agent
instructed to match on @Product category and @Required certifications
links the closest-matching supplier record.
Reference record data
Type @ in Instructions to insert a stable reference to another field. Mark a
reference required when generation should wait for a value, or
optional when the agent may run without it. Renaming the referenced field
does not break the instruction.
Conditional references branch the instructions on another field's current
value — is present, is empty, equals, or not equals — each branch
with its own nested instructions. Use this to change what an agent writes, or
whether it writes anything at all, based on the record's state.
Example: "If @Stock level is empty, write 'Contact us for availability';
otherwise summarize @Description."
Run modes
Every field referenced by the instructions, an image/video source, or Deep Match instructions is automatically tracked as a source field. Field Agents default to manual generation — nothing runs until you trigger it. Switch a field to automatic and it regenerates whenever one of its source fields changes.
Image and video agents require a one-time approval before automatic mode can be turned on, since source-field changes can consume credits and replace agent-generated media on their own — human uploads are never overwritten by automatic runs.
Run and review generations
Run one cell, selected records, the current filtered view, or the full collection. Before charging credits, Cheqq shows a ten-minute quote with the exact eligible cells, protected human values, and estimated cost. Large collections run in deterministic batches of up to 10,000 cells.
Cells show never run, queued, running, provider waiting, ready, outdated, failed, blocked, cancelled, and human-edited states. Open a completed cell to inspect its inputs, model, cost, result, web citations, and prior generations.
A human edit takes ownership of a cell and pauses automatic updates until you explicitly resume them.
Web sources
Web access is off by default and must be allowed by workspace policy. When it is enabled, the active result shows a source-count badge. Open it to review the title, domain, and external link for each citation; citations remain attached to the immutable run rather than being mixed into the typed cell value.