Modular AI-Agents

A step-by-step flow makes a promise about how often this happens: once, in order, and then you are done. Agents are not like that. An owner comes back to soften a fallback message, swap the model, add a document, narrow which courses the agent is allowed to read. Under the flow, every one of those visits meant walking four screens to change one field.

Two of the four steps were also named after nothing in particular. General Settings and Conversation Settings are what a group gets called before anybody has decided what belongs in it, and things arrive in them by default rather than by argument. The fourth step gives it away: the opening message and the fallback message sat together because both had the word conversation in the title. They went to different tabs the moment anyone asked what an owner is actually doing when they write each one.

What shipped is three tabs and no sequence at all. Design is how the agent presents itself: name, colour, the first thing it says. Prompt is how it behaves: model, mode, personality, and what it does when it does not know. Knowledge is what it may draw on: courses, documents, member properties, and the limits on each. There is no more forced step-by-step process.

Two constraints held the whole way through. It had to feel like the rest of the product, which is premium and quiet. And it had to be explainable in one sentence by somebody in sales who is not a designer. Both of those pull hard against elaborate interfaces, which turns out to be a useful pressure to design under.

Then I designed the page that has to explain it. I was hired as a UX/UI designer. However, when the opportunity arose to manage the websites as well, I took it, so alongside the product work I now design and run both company sites. LearningSuite Intelligence is mine end to end: I designed it, briefed the graphic designer on the visuals, built it in Webflow and took sign-off from the owners directly. It is also the largest page I have ever built, and this only after having worked with Webflow for about three months.

Explaining a configurator is its own problem. Nobody scrolling has an agent in front of them, so the page has to make an abstract capability feel concrete before it asks for anything.

I was learning as I went, and that was noticeable in the process. I set the main heading as an SVG because it looked exactly right, which meant the most important line on the page was a picture of words instead of words: nothing for search to read, nothing for a screen reader to say. I also designed it on a fast internet connection and almost forgot about users on a slow one, so optimising the whole page for weight became a pass at the end rather than a constraint I worked under from the start. Both are the kind of thing you only make once.

The larger thing I was learning is what intentional design means on a page rather than on an object. I know how to make a physical thing declare what it does; doing that with scroll, rhythm and restraint was new, and I was doing it on the biggest thing I had made with a tool I barely knew. Webflow also limits you with its own UI, so I iterated with Claude on the custom code that got past those limits, which let me try more versions than I could have built with Webflow alone. I specified, directed and reviewed.

The webpage is the most developed one I have made at LearningSuite, and because it is public it is the one piece of my production work anybody can open and judge for themselves, mistakes included.