Agent input_data

Configure the agent job_type with an exact prompt and registered tool objects, including mandatory input_data tool rules.

Purpose

An Agent is a reusable, prompt-driven workflow step that can inspect, research, and update product data through allowed tools. Create and edit it on Agents, not Tasks, and set job_type to agent.

Valid JSON example

{
  "prompt": "Review the product information, assign the most suitable Shopify category, and preserve claims that the source does not support.",
  "tools": [
    { "name": "read_input" },
    { "name": "write_output" },
    { "name": "list_fields" },
    { "name": "search_shopify_taxonomy" }
  ]
}

JSON fields

Field	Type	Required	Values and meaning
`prompt`	string	Yes	Non-empty instructions describing the outcome and constraints.
`tools`	object array	Yes in the editor	Each item has a non-empty registered `name`. Use the current Tools tab as the authoritative available list.

Mandatory tools

The Agent editor always enables these tools and does not allow them to be switched off:

Tool	Purpose
`read_input`	Read the product details available to the workflow step.
`write_output`	Write the product result that later workflow steps receive.
`list_fields`	Discover which product fields the agent can read or update.

The editor also currently offers optional tools:

search_shopify_taxonomy finds a suitable Shopify category.
search searches the web within the per-run limit.
scrape reads a web page within the per-run limit.
reverse_image_search finds sites or similar results for a product image within the per-run limit.

New Agents start with taxonomy search, web search, and page reading enabled; reverse-image search starts disabled. Optional selections can be changed in the editor. Tool availability can evolve, so never invent a name or copy one from a job payload without confirming it in Tools.

Validation and run-time behavior

The prompt must contain non-whitespace text.
Every tools item must be an object with a registered, non-empty name.
The editor re-adds all mandatory tools when it saves, including after a draft change made through the Assistant.
A prompt-only low-level payload may be accepted by some programmatic paths, but it is not the supported editor contract. Keep the mandatory tools in authored input_data.
Search and page-reading tools bring external information into the run. Tell the Agent what evidence to prefer and what it must not infer.
Never put API keys, store credentials, webhook secrets, or signed URLs in the prompt or tool objects.

UI guide

Use Agents and Agents in workflows for the corresponding editor and workflow behavior.