Installing an agent creates your own configured copy of a library agent. You reach the wizard from Agents ▸ Library: Install agent on a card's detail drawer, or Configure on one you already run.

Installing does not make the agent run. Agents run only when a hook or a scheduled task invokes them, which is what the second step of the wizard helps you set up.

Step 1: Configure

Use as-is, or fork to custom

The first choice is how much of the agent you want to own:

  • Use as-is (recommended): you set the agent's parameters and nothing else. The prompt stays owned by Core dna, so the install keeps improving as the library agent improves.
  • Fork to custom: you get the agent's prompt as a starting point and edit it directly. A fork keeps its own copy and no longer receives library updates.

Fork when the agent is nearly right but needs to know something specific about your business that no parameter covers: your house style, a product naming convention, a policy it must respect. Do not fork just to change a value a parameter already controls.

A fork cannot widen what the agent can do. It inherits the same tool list and the same writable fields as the library agent, and neither can be edited. Forking changes the agent's instructions and judgement, not its reach.

Forking exposes three extra fields: System prompt, Model (an optional pin to one specific model; leave blank to keep using the tier) and Max iterations.

Parameters

Each agent declares its own settings: how far back a report should look, how many suggestions to cap a page at, what tone to write in. They are described on the form and listed per agent in the Library agent reference.

Model tier

Recommended, or an explicit Small / Medium / Large. Leave it on Recommended unless a specific agent is underperforming (try one tier up) or costing more than it is worth (try one tier down).

LLM provider

Client default, or a specific provider. Only providers that already have a saved, enabled key appear in the list.

If the list is empty, no provider key has been saved for your client yet; see Providers & budgets. That screen requires a superadmin user, so if you are an admin without superadmin rights you will need one to add the key before you can finish here.

Setting a provider on the install overrides the client default for this agent only. That is useful when one agent does better on a particular provider, or when you want a costly agent billed to a different account.

Action mode

How much the agent may do without a person:

  • Read-only (recommendations only): proposes; never writes.
  • Accept to action: proposals wait in the Recommendations inbox for a person.
  • Auto-apply with veto window: proposals apply on their own after a delay, unless rejected first.

Choosing auto-apply reveals two more settings:

  • Veto window (hours): how long a proposal sits before it applies. Set this to comfortably longer than the gap between someone actually looking at the inbox. A two-hour window on a queue reviewed once a day is auto-apply with extra steps.
  • Auto-apply confidence threshold: between 0 and 1. Proposals the agent is less confident about than this stay pending for a person no matter how long the window is.

Start read-only. Move an agent to a more autonomous mode once you have watched a few weeks of its recommendations and agree with them.

Escalation

Where the agent turns when it needs a human decision, and where its reports go:

  • Escalation group and Escalation user: who is notified in the admin panel. If both are set, only the user is notified.
  • Escalation chat provider and Escalation chat channel: an optional Slack channel. When one is configured it is always notified in addition to the admin notification.

A chat channel is also how report-style agents deliver: the Sales Report agent posts its summary to the configured channel. Setting up Slack is covered in Chat channels.

The summary and the preview

The right-hand rail summarises what you are about to create: mode, whether library updates apply, action mode, provider, and an estimated cost per run with the model and token basis behind it.

Below it, Live preview runs the agent once against an input you type, so you can see how it behaves before committing. The preview:

  • runs on your API key and costs real money;
  • calls no tools, so give it the content you want it to work on rather than expecting it to go and find something;
  • cannot escalate, and saves nothing: no invocation, no recommendations.

It returns the agent's Decision, its Proposed recommendations and an Outcome. Use it to check tone and judgement, especially after forking a prompt.

Step 2: Create a hook

After the install saves, the wizard offers to create the hook that will run it. Pick a trigger type and choose Create hook, or Skip for now if you would rather wire it up yourself.

The hook it creates is unpublished and asynchronous, with the AI Agent action already pointed at your new install. You finish it in the hook editor:

  • Web event triggers are created ready to use.
  • Entity event and Scheduled event triggers need you to choose the exact trigger: which entity and which method, or the schedule.

The hook does nothing until you publish it. This is deliberate: it gives you a chance to check the trigger and the agent input before the agent starts running against live content.

For what the AI Agent action does and the values it leaves for later actions, see Automations ▸ Hooks.

Editing an install later

Configure on the library card reopens this wizard. Changing settings affects future runs only; recommendations already in the inbox keep the terms they were created under.

Switching an existing install from Use as-is to Fork to custom copies the current library prompt as your starting point, and stops future library updates from reaching it.