/
Blog
/
How we run a Webflow site with AI agents (and where a person still signs off)

How we run a Webflow site with AI agents (and where a person still signs off)

We run our own site with an AI agent every working day. The setup, one real change, and the rules that keep a person in charge of Publish.

Published on
November 16, 2026
Read time
A changeset card for changing the H1 on all 17 case studies, approved with one condition and published to webflow.io, with the Publish button left to a person
Summarize with AI

Key takeaways

  • Four parts, one of them new: the Webflow project, Webflow MCP, a webflow.io staging site, and a person who publishes the custom domain.
  • Every change follows one loop: the agent reads, posts a changeset, waits for approval, builds on staging, verifies, and logs what it did.
  • Rules make it safe: staging-only publishing, a snapshot before any code or schema edit, native Webflow first, and a stop on anything that can't be undone.
  • People keep design, voice and Publish. The agent wins on wide, repetitive work: site-wide sweeps, bulk CMS edits and structured data.

We run our own website with an AI agent every working day. The agent reads the Webflow project, proposes a change, builds it on a staging copy and checks it. A person approves the change before it starts and publishes it when it's done.

By mid-November 2026 that loop had shipped 158 logged rounds of changes to stan.vision: copy, CMS fields, schema, templates, motion, a full SEO and AEO pass. Nothing went live without a human click on Publish.

This article shows how that works in practice. I'll walk through the setup, one real change from start to finish, the rules that keep it safe, and the places where the agent still hands the job back to us. If you want the definition first, start with What is an agentic website.

The setup is four parts and only one of them is new: the Webflow project, Webflow MCP, a webflow.io staging site and a person who publishes

The setup: four parts

The whole system is four pieces, and only one of them is new.

  1. The Webflow project. Our live site, with its components, variables, classes and CMS collections. The agent works on this project, not on an export or a copy.
  2. Webflow MCP. Webflow's official connector lets an AI agent read and edit pages, components, styles, CMS items, schema and the sitemap. It's the new piece. We wrote about it in Webflow MCP: what it is and how we build with it.
  3. A staging site. Every Webflow site has a free webflow.io subdomain. The agent publishes there and nowhere else, so every change can be checked on a real URL before customers see it.
  4. A person who publishes. I review staging and push the custom domain live myself. The agent never has that step.

Around those four sits a written rulebook: how we name classes, where custom code may live, what needs a snapshot first. The agent reads it at the start of every session, the same way a new developer would read our handbook.

One change, start to finish

On 16 November 2026 we changed the H1 on all 17 case studies. Before, each page's H1 was just the client's name, "Rillet", which tells a search engine almost nothing. Here is how the agent handled it.

  1. It read before proposing. The agent pulled the case study template, found the H1 bound to the CMS Name field, and noticed the intro paragraph below it was marked up as an H2.
  2. It posted a changeset. One message: add a plain-text H1 field to the Works collection, write 17 H1s from each page's meta title and first result, rebind the template heading, and change the intro to a paragraph. It listed all 17 lines, for example "Rillet: a Webflow Enterprise site for a $1B ERP".
  3. I approved, with one condition. The "Up next" scroll transition at the bottom of every case study might depend on the heading, so I asked the agent to make sure it wouldn't break.
  4. It checked the risk first. The agent read the transition script on staging and confirmed it takes the next project's name from the CMS, not from the H1. Only then did it make the change.
  5. It built and verified on staging. New field, 17 values, template rebound, published to webflow.io. It then fetched all 17 pages and confirmed each had exactly one H1 with the new text, and scrolled one case study to the end to watch the transition load the next page.
  6. It reported what it couldn't do. The intro paragraph stayed an H2: the component's tag setting only accepts h1 to h6. The agent said so and offered a native alternative instead of patching it with custom code.

My part was two messages: the approval with its one condition, and a read of the 17 headlines.

The six rules the agent follows on every change, from changeset first to stop when an action cannot be undone

The rules that make it safe

An agent with edit access to a live site is only as safe as the rules around it. These are ours, and the agent follows them on every change.

Rule Why it exists
Changeset first, then approval, then the changeNo surprise edits. I see what will change and where before anything moves.
Publish to staging onlyThe agent can't take anything live. Only a person publishes the custom domain.
Snapshot before touching code, tokens or schemaEvery embed, variable and JSON-LD block is saved to our project log first, so any change can be rolled back exactly.
Native Webflow firstClasses, variables and components before custom CSS. Code goes in an embed inside the component it belongs to.
Log every roundEach change ends with a note: what changed, where, how it was checked, what's still open.
Stop when an action can't be undoneUnpublishing a CMS item removes it from the live site at once, so the agent waits for the next human publish instead.

The last rule came from experience. Setting a published item to draft doesn't take its page down, and the only API call that does acts on the live site immediately. The agent spotted that, held the step, and added it to my publish checklist.

The StanVision team reviewing work at the table: design and voice, Designer-only steps, rich text with images and publishing stay with a person

What agents do well, and what still needs people

The agent is fastest at work that is wide and repetitive. It is weakest at taste and at the few corners of Webflow the API can't reach.

Where the agent wins

  • Site-wide sweeps. In one pass it fetched all 145 pages in our sitemap and checked every mention of our partner tier and award count against our profiles. A person would sample ten pages and hope.
  • Bulk CMS work. 17 new H1s, 87 journal articles retagged by topic, FAQ answers kept in sync with their schema.
  • Structured data. JSON-LD on every page and template, written through Webflow's native schema field rather than custom embeds.
  • Checking its own work. After each change it opens the staging pages, compares them with the snapshot and reports what it saw.

Where a person still decides

  • Design and voice. The agent proposes. On our Work page I rejected several grid layouts before one felt right. That judgment stays with us.
  • Designer-only steps. Some things only work inside the Webflow Designer, such as binding a text to a CMS item's built-in Created On date. The agent does these through a browser, slowly, or hands them to us.
  • Rich text with images. Reading a rich text field through the API and writing it back swaps the optimised images for the originals. The agent edits those in the CMS editor instead.
  • Publishing. Always a person.

How to start on your own site

Start small and keep publishing in human hands.

  1. Connect Webflow MCP to the AI tool your team already uses, with access to one site.
  2. Write the rules down first. Naming, where code lives, what needs approval. Use our starter rules below as a base.
  3. Give the agent a read-only job. An SEO or content audit of your own site shows you how it reasons before it edits anything.
  4. Move to staging edits. Let it change CMS items or meta data and publish to webflow.io only. Review there, publish yourself.
  5. Keep a log. One note per change. It's your rollback plan and the agent's memory between sessions.

Starter rules for an agent on your Webflow site

These twelve come from the rulebook our agents follow on every client build. Paste them into your agent's instructions and adjust the names to your site.

Safety

  1. Publish to the webflow.io staging subdomain only. A person publishes the custom domain.
  2. Before any change, post a short changeset: what changes, on which pages. Wait for a yes.
  3. Save the current version of any embed, custom code, variable or schema block before editing it.
  4. Write a log note after every change: what changed, how it was checked, what's still open.

Building

  1. Reuse existing components, classes and variables first. Anything new needs a one-line reason.
  2. Anything used more than once becomes a component. Same structure with a different look is one component with variants.
  3. Style with Webflow classes and variables. Custom CSS only for what the Style panel can't set, and say which property.
  4. Code for one section lives in an embed inside that section's component, so it moves with it.
  5. No hardcoded colours in code. Use variables, so light and dark modes both work.
  6. Pin every external library to an exact version, and leave no console.log in production.

Content and SEO

  1. One H1 per page. Every page gets a meta title and description, every image alt text.
  2. Structured data goes in Webflow's page schema field, not in custom code embeds.

If you'd rather have a team set this up and run it with you, that's our agentic Webflow websites service. On Webflow Enterprise, the same loop adds roles and staging branches; see Webflow Enterprise.

Frequently asked questions

Can an AI agent break my live Webflow site?

Only if it can publish. Ours publishes to the webflow.io staging subdomain only, and a person pushes the custom domain live.

Do I need Webflow Enterprise to work with agents?

No. Our own site runs this way on Webflow's CMS site plan. Enterprise adds roles, page branches and governance, which matter more as more people and agents edit the same site.

What can the agent not do in Webflow?

A few Designer-only actions, such as binding a text to an item's built-in Created On date, and some component settings the API restricts. It hands those back or does them through a browser.

How do you undo an agent's change?

From the snapshot it saved before the change. Every embed, variable and schema block is copied to our project log first, so a rollback restores the exact previous version.

Stan Kirilov

Founder & Experience Director

Stan has led design for over 20 years, first as a hands-on designer, now as Founder & Experience Director at StanVision. He works with SaaS and fintech companies, from first-time founders to global brands, on products and websites built to convert, not just to look good.

He serves on the Awwwards jury and still reviews every project that ships. His standard is simple: if a design doesn't perform, it isn't finished.

Webflow
AI design

What is an agentic website? A practical definition for marketing teams

Nov 12
9
min read
Stan Kirilov
Stan Kirilov
Founder & Experience Director
Homepages of the ten best Shopify agencies for 2027 laid out as a tilted wall of screens
Web design

Best Shopify agencies for 2027, compared by expertise and results

Nov 5
11
min read
Stan Kirilov
Stan Kirilov
Founder & Experience Director
Webflow

WordPress to Webflow migration: how to move your site and keep your rankings

Oct 27
8
min read
Stan Kirilov
Stan Kirilov
Founder & Experience Director
Webflow
AI design

Webflow MCP: what it is, and how we actually build with it

Oct 26
14
min read
Stan Kirilov
Stan Kirilov
Founder & Experience Director

Your website called. It wants to talk.

One call. Tell us what the new site has to do, and you leave with a real scope and a timeline. No deck, no "circle back".