An AI agent that runs the website
Texts, images, pages, publishing: the site you are reading is updated by an AIDeskPro agent. How it is built, what it does and why there is no WordPress any more.
The case
Two sites, one agent each. opengate.biz is Open Gate’s website: 299 pages in Italian and English, a blog with content since 2014, contact forms, product landing pages. Until October 2026 it was WordPress, with its dashboard, plugins and weekly updates. On 9 October an agent rebuilt it as a static site in one working day, with a person re-reading the pages: same URLs, same content, same forms. Since then an AIDeskPro agent runs it.
aideskpro.com, the site you are reading, was born this way: 65 pages in two languages, one blog article a week, the agent and tutorial pages. Here too every change goes through the agent.
Both sites are git repositories on Google Cloud. The generator is Hugo, the hosting is a CDN. There is no database, no server-side language, no back office to learn: there is a Console where you talk to the agent and a staging copy of the site where you look at the result before publishing.
The agent
It is called CMS and it is an AIDeskPro Configuration. Unlike the agent that writes sales proposals , it has no document indexes: its resource is the Workspace, where it keeps the site’s repository, an operating guide and three scripts.
- The guide (a text file, AGENTS.md) says where things are, how a page is built, which components exist, what must never be touched without explicit instructions (page URLs, redirects) and the rule above all rules: staging first, production only on request.
- setup.sh reinstalls the tools (Hugo, the hosting CLI, the repository credentials) every time the sandbox restarts.
- deploy.sh builds the site and publishes it to the chosen environment, staging or production.
- check.sh checks the build and the internal links before every deploy.
A CMS execution never ends: it is paused when not needed and resumed when there is something to do, with the whole history in the Console. With image generation enabled, the agent also produces the visuals in the site’s style.
One deploy, step by step
This page and the one on the sales agent went online like this. The texts were already written and committed to the repository; the agent did the rest.
1. The request
You resume the paused execution and write in the Console:
Run setup.sh. Then pull master, build and deploy to staging. Stop before production.
2. The setup
The sandbox had been off since the day before and restarts with no tools. The agent runs the script, which reinstalls Hugo, the hosting CLI and the link to the repository. One minute. The Console shows every command and its result.
3. Pull, build, check
The agent pulls the changes, builds the site (the two new pages, the updated cards, the sitemap) and runs the link check. If an internal link points to a page that does not exist, it stops here and says so.
4. Staging
The deploy goes to the staging environment: same site, different address, not indexed by search engines. The agent replies with the link and stops, as asked. A person opens the pages, reads, tries the links and the PDF.
5. Production
“Publish.” The agent runs the same script towards production and in under a minute the pages are on the CDN. If it touched a file in the meantime, it commits it and describes the change.
When the agent writes the text
On opengate.biz the job starts earlier. On 14 October 2026 marketing asked in chat for a post explaining the line at the bottom of the site, “Website managed by an AIDeskPro agent”, in an informal tone. The agent wrote the text in Italian and English, generated the featured image with the yellow frame of the others, created the two pages and put them on staging.
Then the corrections, one at a time, as with a colleague: add a section on what AIDeskPro is and one on WordPress; in the image remove the title, centre the graphic and put the logo bottom left; in the first example, instead of the email-signature logo, tell the story of the old site’s clone. Four commits in one afternoon, each with its description, and the post went to production with an “ok”.
In the same weeks the agent recreated the missing images of the posts from 2015 to 2019, aligned the English AIDeskPro page with the Italian one, videos included, published the post on the FARE seminar in Bari with the video and, on 3 November, updated the contact-form function with an LLM-based antispam filter. All asked in chat, all through staging.
The result
opengate.biz. 299 pages in two languages, content since 2014, migrated from WordPress in one working day with the same URLs as before. From 9 October to 3 November 2026 the agent signed twelve commits: new posts, regenerated images, aligned pages, the form function. PageSpeed scores of 97 on mobile and 99 on desktop, measured on the emulator.
aideskpro.com. 65 pages in two languages and one article a week since June 2026, 210 commits. The agent pages, the tutorials with screenshots, the pages you are reading.
What is gone. The admin dashboard, the plugins to update, the database to protect, the three people for one post. What remains is a person who decides what to publish and looks at it on staging before saying yes.
On both sites the agent uses AIDeskPro’s default model, Claude Sonnet 5. The model can be changed from the execution page, even while the work is in progress.
What PHP2Shell has to do with it
In October 2026 two critical flaws in the PHP engine, the language behind nearly 70% of websites, were made public: CVE-2026-109618 and CVE-2026-109619, known together as PHP2Shell. They allow code execution on the server without credentials, had been there for years and were found by researchers assisted by AI models, in weeks instead of months. The first exploitation attempt was logged on 17 October, four days before the patch of 21 October. Mass attacks followed within hours.
Anyone with a WordPress site spent that week checking versions and updating. opengate.biz, migrated eight days earlier, had nothing to update: no PHP, no database, no exposed admin panel. Just HTML files on a CDN.
The same speed an agent uses to build a site, someone else uses to break in. A static site run by an agent removes the attack surface instead of defending it.
