Skip to main content
An agent column writes a prompt for one column on one table. Save that work on Agents and you can add the same agent to any table.
The Agents page with two sections: a row of template cards under Start with a template for Company research, ICP fit, How they'd use it and Email message; and a Browse agents section below with a search box, a New agent button, and one saved agent named Company research agent.

The Agents page.

​
Build and save an agent

Click Agents in the sidebar, then pick a card under Start with a template. See all templates opens the full gallery, and New agent opens a blank editor. A template fills in the name, prompt, inputs and outputs. Adjust the Effort, Inputs, Outputs and Tools cards, then click Save. The agent then appears under Browse agents.

​
Deploy it to a table

Click Add to table in the editor header and choose a table. Point each input at a column, or click Match inputs by name to fill in the ones it can match. Name the column last. Its split button can add the column and run a first batch at the same time.
Adding and running rows spends credits at once. Add and don’t run is the free option.

​
Good to know

A row’s menu has Edit agent, Duplicate agent, Show where it’s used and Delete agent. The star marks a favourite.Show where it’s used lists every table and column the agent is deployed on, and jumps to them.
Duplicate agent makes a private copy for you, reset to version 1 with no history.Save as template, in the editor’s menu, publishes a copy into your team’s gallery under a name and description you give it. Both are copies: editing the original later changes neither.
Each deployment maps its own inputs, so remapping on one table leaves the others alone. Use the whole row is offered only when the prompt declares no inputs of its own.
When the outputs declare more than one field, the deploy dialog names a parent column and lets you rename or skip each field. It refuses a name that collides with a column already on that table.
Train agent, beside Save, runs the agent over sample rows so you can judge each result. Approving a revision saves a new version. The button needs a name, a prompt and at least one input, and it saves your edits before training starts.
The version chip beside the name opens the history, and any earlier version can be reverted to.
Its prompt and outputs become read-only on the column, and a card at the top of its editor links back here. That behaviour is covered on Agent columns.

​
Next steps

Build an agent

Start from a template or a blank prompt, then write the prompt.

What agents cost

See what moves the price of an agent run.