What an AI agent needs from your CMS

2 min

The BetterCMS team Engineering

Summarize with AI

TL;DR, key takeaways

  • An agent needs to read the schema before it can write anything useful.
  • Give it scoped tools rather than a human login.
  • It should propose changes for a person to approve, not apply them.
  • Every change an agent makes belongs in an audit trail.

The current way teams use a model on their content is to paste a page into a chat window and paste the answer back. It works, it does not scale, and it loses every link between the words and the fields they came from.

Connecting an agent to the CMS properly is a different problem, and most of it is not about the model at all.

It needs to read the schema

An agent asked to update a product page has to know that a product page has a summary, a set of plans and a reference to a case study. Without the schema it guesses, and a guess in a content system is a field written into the wrong place.

A machine readable model is the single biggest thing that determines whether an agent is useful here. It is also, conveniently, the same thing that makes the content reusable for humans.

JSON

GET /schema/productPage

{
  "_type": "productPage",
  "fields": [
    { "name": "summary",   "type": "text",   "required": true },
    { "name": "plans",     "type": "array",  "of": "pricingPlan" },
    { "name": "caseStudy", "type": "ref",    "to": "caseStudy" }
  ],
  "permissions": { "write": ["editor", "agent:proposal"] }
}

It needs scoped tools, not a login

Giving an agent a session and letting it drive the interface is the worst version of this. Every action it can take is every action a person can take, including the destructive ones, and nothing about the interaction is inspectable afterwards.

Named tools are the alternative. Read an entry. Propose a patch. List the pages in a collection. Each one has a signature, a scope and a permission, so what the agent can do is a list somebody wrote down rather than a consequence of what it happened to click.

It needs to propose, not apply

The useful unit of agent work in a CMS is a proposed change set. Twelve entries, a patch for each, shown before anything is written. A person reads the diff, approves nine, discards three.

This is not a limitation to be removed once the model gets better. Review is where the trust comes from, and a change set that was reviewed is one somebody is prepared to defend.

  • Every change is shown as a field level diff before it is applied.
  • Approval is per document, not all or nothing.
  • Everything applied is reversible.
  • Publishing stays a human action.

It needs to know your rules

A general model writes general copy. It will use words your style guide bans and it will describe features the way a competitor does, because that is what it read.

Whatever the mechanism is called, the agent needs your terminology, your tone and your product facts in front of it every time it writes. That is a content problem, and it belongs in the CMS next to the content it governs.

An agent with your schema and your rules is a colleague. An agent with your password is a liability.

It needs an audit trail

When a page says something wrong, somebody has to find out how it got there. A run that records what was asked, which tools were called, what changed and who approved it answers that in a minute. Without it the answer is a shrug, and the next decision is to turn the agent off.

More articles

View all

Your next website starts here

See BetterCMS running on your pages, not on a sample site.