Thomas Doganay
Thomas Doganay

AI Design Engineer

← Back to projects

Concept · 2026

Understudy

A landing page concept for a browser agent that learns a chore by watching you do it once, then does it for you from then on.

Role
Product Design / Landing page
Year
2026
Stack
Figma, Copywriting, Landing page

Explaining a new kind of agent in one scroll

Understudy is a concept for a browser agent you teach by example. You do a chore once while it watches, and from then on it does it for you: invoices, bookings, reorders. I designed the landing page to explain a new kind of product in under a minute, without a single technical word.

The design problem is that nobody has a mental model for an agent you teach by example, so the page cannot lean on a category. My approach was to put the working interaction at the top, a single input and one button, and to spend the rest of the scroll answering the three questions a sceptical visitor asks in order: how does it learn, what happens when things change, and what stops it doing something I did not want.

How it should feel

Bright, friendly and a little bit magic, but never vague. The blue gradient carries the optimism; the orange “Agent” cursor makes the agent visible, so you always see what it is doing on your behalf.

Every claim on the page is shown, not told. Orange is reserved for the agent and for the one action, Teach it, so the eye learns that orange means “this is where something happens”. Everything around it stays white, warm grey and near-black, which keeps the page quiet enough for the product demos to be the loudest thing on screen.

Up close

The promise and the first step, side by side.
Real errands as real cards, each ending with “asks first”.
01

Show. Watch. Repeat.

A visitor who has just read the promise now wants to know how, so I answer with three numbered cards and a subline that sets expectations: three steps, one demo, no scripts. Each card is a tiny product moment, not an icon: an input with an orange Record button, a checklist with a progress bar that stops at Picks last month, and a counter reading 12 steps learned. The third card shows its boundaries as well, with Runs every Monday and Adapts to layout changes in green and Asks before paying in red-orange, so the one limit stands out. The row of assistant logos above says it works next to tools people already use.

02

When a page changes, it finds the new way

The sceptical question here is whether this is just another brittle script, so I show the exact failure that breaks scripts and how this survives it. The headline and two short lines sit left with a single Teach it button, and the demo fills the right half. Three restaurant cards sit side by side with the middle one labelled New layout, carrying a Get 10% off pop-up, and CHANGED under it, while the outer two read LEARNED. The orange Agent cursor lands on the card that matches, and a confirmation below, Booked at Trattoria Nove, says it closed the pop-up and carried on. Small uppercase status labels carry the story without a paragraph.

03

It remembers how you like it

After reliability, the next worry is whether the agent knows me, so this section is one centred headline over a single exchange. Two panels sit side by side and are labelled You said and Understudy, so it is clear who is who. On the left, a short request, Reorder the Norra jacket in my size, gets a one-line answer, Done. Size M, as last time. On the right, the shop shows what the agent did: S sold out, M highlighted in orange with 2 left in green, L in stock. Showing the sold-out and stock states makes the choice of M feel like judgement and not coincidence, and the small caps eyebrow, Learns from you, names the idea before the headline does.

04

Your errands, off your plate

This is the section for the visitor who is nearly convinced and wants to know what exactly it will do for them. I used a two by two grid of equal cards, each with a UI fragment on top and a short title and one-line caption below, so it reads like a quick catalogue of errands and not a feature list. The cards go from low risk to high: it checks opening hours first, books by asking, finds a button even when its label changes, and finally pays only with your OK. Only the last one carries a price, 13.50 euro, and the Agent cursor sits right on the Pay button, which is where trust is decided. The orange outline marks what the agent is focused on in each card.

05

Teach it one chore

By the end of the scroll the visitor has either decided or left, so the closing block does one thing: it repeats the same input and the same orange Teach it button as the hero, so there is nothing new to learn at the moment of action. The headline names the benefit in plain life terms, Get your evenings back, and the line under it reduces the effort to about a minute. On the deep blue gradient, the white input is the brightest object, which pulls the eye straight to it. The four small chips floating around the edges, such as Invoices: 12 and Reorders: 5, hint at the chores other people have taught, without competing with the headline.

See the full pageThe whole design, top to bottom, in full quality.
Open the original file ↗

Looking back

Explaining an agent is mostly about restraint: what it does, where it stops, and when it asks. This page is my answer to that, in one scroll. It is a concept, so none of the demos are real runs. If it moved forward I would test whether the opening input works as a starting point for first-time visitors, and I would check that the claim of a minute is one the product can keep.

Next projectTidemark →