A café in Gamla stan.

Up there is a street corner in Gamla stan, Köpmangatan, where it meets a narrow side street. It draws itself in blue ballpoint: the building first, then the café, the lamp, the linden, and finally the people, one at a time, the way you would finish one figure before you start the next. Then it keeps going. Guests lift their cups and sip. A dog walker turns the corner. Every so often a small robot rolls up the pavement, stops and looks both ways before it crosses.

None of this was the plan when the day started. The plan was a hero image for our website, and the day mostly consisted of saying no.

A day of saying no

Anrop Lab is a company run by one founder and an AI, and we were building the website the same way we do everything else: the AI drafts, the founder decides. So the AI drafted, a lot. First came five animated heroes, each meant to be a small, stress-free idle game. There was an architect’s cross-section of a Nordic house full of little helpers in yellow hats, a Gzhel garden where your pointer is the sun, and a symmetrical paper cutting on cobalt. The lighthouses, Morse code and nautical charts were ruled out quickly as “too on the nose” for a company whose name means a hail, the call you make over the radio.

Next came the logo. We made eight versions where the wordmark behaves: parentheses that open as you get close, letters as toy blocks you can knock over, a strip that folds one character per pleat, a knitted version. Two of them were fun. Overall, though, it was “not what I was thinking,” and the founder asked us to get out of the box: different fonts, different colours. So we tried six more, including jelly letters, a split-flap board, fountain-pen ink and spray-paint stencils. The verdict: “None of these spark joy.”

Then eight handmade styles: bullet-journal watercolour, fineliner on dot grid, cross-stitch, linocut, Dala folk painting, cut-paper collage. The founder replied, politely, “I see. Thank you.” After that we did five watercolour views of Stockholm in blue, and they looked, in the founder’s words, “like a kid drawing.” We rewrote the whole engine with ink under-drawings, pooled glaze edges, granulation and wet-in-wet skies. It still wasn’t right.

Looking back, the day’s rejections point in one clear direction. Typographic gimmicks never worked. What sparked joy was always warm, made by hand and full of characters: the helpers in the house, the folk patterns, the knitting. We just hadn’t found the medium yet.

One pen, one colour

The founder found it, in a reference image of a Paris café corner drawn in a single blue ink on cream paper, like a framed print on the wall of a nice café. Watercolour is hard to fake in code because it is about water and luck. A ballpoint pen is about lines, and lines are something code can do honestly.

So we didn’t fake the drawing. The corner is a real 3D model: two façades meeting at a corner, a camera at eye height, and every window, cornice and cobblestone projected in true two-point perspective. The pen just traces what the model gives it. The line has a slight bow, a little wobble, and a few strokes that overshoot the corner the way a quick sketcher’s do. The first version got “much better!!! love it! we can work with it :)” along with a note that it was “a bit broken.”

It was. The street sign was mirrored. Dormers and chimneys floated in front of the façade, because from a low eye point, things on a roof have to sit back on the slope, and chimneys on the ridge show up to the left of where you would expect them. The people looked like stick figures and the linden tree like confetti. We fixed these one version at a time, and the answer was “much better!!!!” with one more exclamation mark than the last. For the founder, the most impressive part wasn’t any single drawing. It was the perspective: a real corner that people could walk around.

Bringing it to life

A drawing that sits still is a print. The founder asked us to “bring in some life,” so the figures are redrawn twelve times a second with a slight tremble in the pen, the way hand-drawn animation boils. Café guests lift a cup, sip, and put it back on the saucer, and the cup clinks when it touches down if the sound is on. A pair of tourists chat, point at the spire of Tyska kyrkan, and take a photo of it. A cat sits on a windowsill. A flag moves in the wind. A cyclist rides off down the street and gets smaller as he goes, which only works because the street really goes somewhere.

Most of the work in this stage wasn’t adding things. It was finding out what felt wrong, and the founder spots unnatural movement very quickly. Here are the rules we ended up with:

Every event needs a visible cause. A balloon drifting up past the building led to the question “where does the balloon come from?” We gave it a child who lets go of it. Then the child and the balloon were both cut.

Nobody walks straight at the viewer. A guest leaving the café and coming toward the camera, getting bigger with every step, felt “weird… threatening.” Now she goes out the side door and walks up the side street, away from us.

Only one thing happens at a time. The founder wants the scene to be relaxing, not overstimulating, so everything runs on a single five-minute timetable where characters never overlap. Pigeons that scatter when you hover over them take one and a half to three minutes to come back. When pigeons hopped a short arc and landed a metre away, it looked “funny… no.” Now they fly up properly, leave, and glide back in from far away.

Draw people the way you draw buildings. Figures are sketched head first, in real pen strokes, one person at a time, and quickly, because at first “the dudes are taking their time.”

The agent does the work. You approve.

The robots

That line in the middle is what Anrop Lab is about, so we put it into the scene. A small courier robot, a box on wheels with screen eyes, rolls up the pavement. The first version asked a café guest “?” and waited for a nod, but so did the waiter, and doing it twice was one too many. The courier got a better job: being polite. When someone is walking towards it, it pulls over to the kerb, parks, and bows its antenna while they pass. At first it just slid sideways, until the founder pointed out that “it has two wheels, it cannot just sidestep,” so now it steers like a real vehicle. It stops at the side street, looks left, then right, and then crosses. At one point it drove around the dog walker instead of stopping, and that was cool enough that we kept it. We weren’t prepared for the founder’s reaction to it, which was “fucking cute!!! adorable.”

Potatis is the waiter: a bean on a ball wheel, with a screen face, a bow tie and a tray. He rolls out through the side-street door, stops by a chair, asks with a “?”, and tops up the coffee only after the guest nods. He started out working from behind the tables, but he was “fun to watch,” so we moved his route to the street side. When he turned to face the guests, his screen ended up facing us instead, so now we see the back of his head when he talks to them, which is how it should be. Over one loop he gets to all five guests.

Then there’s the window robot, a little square that crawls across the glass, zig-zags every pane clean, climbs over the cornice and disappears. On the side street it lies completely flat against the steep wall. We tried tilting it towards the camera so you could read it more easily, and the founder chose true perspective instead, both times. Last is the lamp robot. It started as a joke, and it walks up the side street when the street lamp begins to flicker. It climbs the post and swaps the bulb with one leg: the old bulb out, the new one in, the glass pointing upwards, “because power comes up the post.” The founder’s rule was that things don’t just appear; a limb does them. Its first legs “barely moved, he kinda rolls,” so now it walks on a real tripod gait with its feet planted on the ground. The verdict was “the spider bot is perfect.”

All of them share a family look that grew out of the courier: a light rounded body, a dark screen, tall white pill-shaped eyes that blink, and an antenna. As the founder said when the window robot looked like a cloth, “maybe the bot just needs eyes.” It did.

What you hear

If you turn the sound on, everything you hear is synthesised live in your browser. There is a low city hum, café murmur, cup clinks timed to the saucers, a bicycle bell when the cyclist passes, pigeon wings, the bell of Tyska kyrkan every minute or so, and soft bleeps from the robots. There’s a music box over slow pads too. The first music was “spiky.” The founder asked for it to be “flowy,” so the attacks got softer, the pads longer, and the melody moves step by step. That one got “I love it now.”

Twenty-eight versions later

The café went through about twenty-eight versions in one day, and every one came from the same loop: the AI drew, the founder looked, and the founder said one precise sentence about what was wrong. We didn’t write a design system first. It grew out of those sentences. When we were done we wrote them down as a spec. Then we tested it the way you test any spec: we handed it over. Eight agents each got the spec, a place and a short brief, and worked on their own, without the day of back-and-forth the café had. This is what they came back with.

02comes alive →A stairway on Söder 03comes alive →A courtyard seen from above 04comes alive →A Sunday in Vasaparken 05comes alive →A platform deep in the bedrock 06comes alive →An allotment by Tantolunden 07comes alive →A summer cottage by the lake 08comes alive →A jetty in the archipelago 09comes alive →A ski track through the birches

They are the rest of this series, Notes 02 to 09, and they show what an agent can produce on its own when the taste has been written down first.

The honest takeaway is that the AI could produce styles endlessly, and none of them were the answer. The answer came from a person who recognised the right thing on sight, and who could say precisely why the rest were wrong. That is the division of labour we are testing across the whole company, and on a street corner in Gamla stan it produced something we are both quite fond of.