An agent-operated website for a dive resort

A resort website and CMS that AI agents can work on through MCP: agents propose edits to pages, products, and dive sessions, staff approve them, and every request is logged with the person who made it.

What was at stake: The resort's website covers its dining, store, dive sessions, and lodging. Staff need to keep it current without technical help, and a wrong price or schedule reaches guests immediately.
The decision: Let AI agents read the site and propose changes, but make every change a preview that a person approves in the CMS before it goes live.
What we chose not to do: No agent can publish on its own; live pages, site settings, and search details change only after a person approves them.

Joh's role: Engineer for the website platform and its AI agent connection

AI Operations Care

Context

The product and operating environment

The resort runs several businesses from one website. The CMS exposes its pages, navigation, products, media, settings, and dive sessions to AI agents through an MCP server, with the same approval rules a member of staff would follow.

Important constraints

Letting non-technical staff update a busy site through an AI assistant without risking wrong prices, menus, or schedules.

Keeping the owner's own wording when an agent updates menu text.

Making sure changes to live pages, settings, and search details always get a human decision.

Knowing afterwards who asked for each change and what it was.

Engineering responsibilities

Website and CMS architecture.

MCP server design: read tools for content, and propose tools for changes.

Preview, confirm, cancel, and undo flow for agent-proposed changes.

Approval in the CMS for live pages, site settings, and SEO, with a link the agent hands to the person.

Request logging tied to the user who made each call.

Technical decisions

Separate reading from changing: agents can look up anything, but every change starts as a proposal with a preview.

Require a person's approval in the CMS for anything that affects live pages, settings, or search.

Log every agent call with the name of the person behind it.

Offer undo as a first-class action rather than relying on backups.

Practical outcomes

Staff can ask an AI assistant to update the site in plain language, and approve the result instead of editing it themselves.

Every agent change is previewed, approved, and attributable, so the owner stays in control of what guests see.

A reusable pattern for connecting AI agents to any business CMS safely.

Technology and systems

MCPCMSApproval workflowsAudit loggingUndo

This case study describes Joh's role and publicly shareable technical scope. It does not expose confidential customer data, private credentials, internal commercial information, or imply an endorsement by the named organization.

Apply the same thinking

The right decision is worth more than the code.

The AI Operations Audit applies the same judgment to your business: what to automate, what to protect, and what to leave alone, in ten business days.