Skip to main content
An agent watches a store for you. It wakes on a schedule or when something happens, reads your store’s data, records what it finds, and, when you let it, acts on what it finds. Three things make an agent:
  • Instructions: what it should do, in your own words.
  • Skills: written procedures it reads when they apply. See Skills.
  • Schedules: when it wakes, and what it does then. See Schedules and triggers.
An agent does nothing until you deploy it to a store. Open Agents in the left rail. The list shows each agent’s mode, how many stores it runs on, how many deployments are running or paused, and how many open findings need you. Filter by Store or by status (Active, Archived or All).

​
Create an agent

Choose New agent, then pick a starting point. A template fills in every field, and everything it fills in stays editable. The wizard has four steps. A finished step collapses to a summary; select its header to change it.
1

Basics

A Name no other agent in the organization has, and an optional one-sentence Description (up to 280 characters) for the agents list.
2

Instructions

What this agent does on top of its skills, in your own words, up to 20,000 characters. Headings, lists and tables all work.
3

Skills

Pick the skills it may read. Glood’s own skills that every agent gets are already ticked and marked Every agent. This step needs the skill:attach permission; without it, the agent gets only the skills every agent gets.
4

Review and create

Lists the checks a template starts with. Create agent saves it.
There is no model or mode to choose. Every agent runs on Glood’s model and starts in Act mode. You can turn a deployment down to Recommend or Observe on its own Settings. See Modes.

​
Write good instructions

The instructions are the agent’s standing brief. The template’s instructions are a good model:
  • Say what to watch and what to do about it, as a short list.
  • Name the skills it should follow, by slug, for example “follow the fraud-policy skill”.
  • Say what deserves a finding and what does not. “Never record a finding for a coincidence” saves you noise.
  • Say what to do when it cannot find a cause.
When you need the agent to call a specific tool, use the tool’s exact name. See Agent tools.

​
Configure an agent

Open an agent and choose the Configuration tab. It is one page of sections; each header summarises its state. Each change to the instructions is kept as a new version, so a past run can always say what it was told. Changing an agent needs the agent:write permission. Without it the Configuration tab is read only.

​
Rules

The Fraud agent’s Rules tune how it scores and flags orders. Each setting shows its default, and Reset puts it back. Only settings you change from the default are saved.

​
Archive an agent

Archive agent, in the Archive section, stops every deployment of it, takes it off the list, and means it can never be deployed again. Its runs, findings and history are kept. Type the agent’s name to confirm.

​
Next

Schedules and triggers

When the agent wakes, and what it does then

Deploy and set the mode

Put it on a store and decide how far it may go

Chat with an agent

Ask it about the store it runs on

Runs, findings and changes

What it did, what it found, and how to undo a change