Rocks, Not Robots: How AI Redesigned Our Logo in 15+ Iterations

Multiple people told me my logo was scary.

They were polite about it. “It’s… distinctive.” “Very memorable.” But the message came through clearly enough: the bracket-eyed grin that was supposed to say “friendly developer” was reading as something far less welcoming at small sizes.

The Logo That Scared People

The CodeWithAgents logo had already been through four iterations during the initial site build — code brackets, text variants, and finally a bracket-glasses nerd face with < and > eyes and a toothy grin. At full size it looked nerdy and fun.

At small sizes — favicons, navbar icons, social previews — it looked like it wanted to eat you. The brackets became angry eyebrows. The teeth became… teeth. The kind that bite. Several people independently pointed this out, always carefully.

I got the message.

Why Stones?

The replacement idea came from an unexpected place: Cairn, the AI system I’d built to persist knowledge across sessions. When I was first building this site with AI, Cairn was already part of the process. The name itself carries the metaphor — a cairn is a pile of stones, each added by a traveler to mark a path for those who come after.

That’s what agentic engineering is at its core. You build systems that accumulate knowledge incrementally, persist across time, and guide the next traveler. Humans have been doing this with stones for tens of thousands of years. We’re doing it with language models now. Same pattern, different material.

So yes — an AI consulting brand with a logo made of rocks. The irony only hit me after I’d chosen the concept. Everyone else in this space has neural networks, circuit boards, or robot heads. We’ve got a pile of stones.

It fits better than any of those.

The SVG Problem

I described the concept to Cairn: five stacked stones, angular and irregular, suggesting the trail-marker meaning. Make it work as SVG at every size from favicon to full page.

Here’s what I learned over the next few hours: AI can reason about visual design remarkably well. It can discuss composition, suggest layouts, understand what “angular” and “organic” mean aesthetically. But when it comes to actually rendering those ideas as pure SVG — path coordinates, polygon vertices, gradient stops — it’s like asking someone to paint a portrait by dictating GPS coordinates.

The challenge isn’t intelligence. It’s the medium. SVG is a set of mathematical instructions that happen to produce shapes. There’s no “make this look more natural” parameter. Every curve is a Bézier control point. Every shadow is a separate polygon with a carefully chosen fill color. The gap between “I understand what good design looks like” and “I can produce the coordinates that create it” is enormous.

We were about to discover exactly how enormous.

Version 1: The Turd

The first attempt was three rounded stones stacked vertically.

I laughed first. Then I groaned. It looked exactly like what you’d find at a spa reception desk — or worse, what you’d find in a field after a large animal walked through. Smooth, organic, symmetrical. Everything a cairn is not.

Real cairns are messy. Irregular. Angular. The stones don’t nestle — they lean on each other with visible gaps. That tension is the whole point. Each stone depends on the ones below it. Remove one and the structure changes.

The 15-Iteration Grind

What followed was a long series of increasingly specific instructions:

Versions 2-5: More stones. Angular shapes. Five stones for five letters in “Cairn.” We got the basic silhouette right — a wide base stone, a small wedge, a medium slab, a tall pillar, and a capstone leaning across the top to form an arch. The shapes were good. But they were flat white outlines on nothing.

Versions 6-8: “Let’s add shading.” This is where things got hard. I wanted three or four emerald shades to suggest volume — light hitting from the upper left, shadows underneath the capstone. The AI understood the concept perfectly. It could explain exactly which faces should be lighter and which should be darker. But translating that understanding into polygon coordinates that actually create the right visual effect? Stone after stone came out wrong. Shadows extending too far. Light areas appearing on the wrong side. Shading stripes that looked identical on every stone despite different orientations.

Versions 9-12: We tried fixing the shading by splitting each stone into light and dark polygons. The approach was sound in theory. In practice, the polygons never quite lined up. White gaps appeared between faces. The more we fixed one stone, the more another one broke. I was adjusting individual vertex coordinates one at a time, and each fix created two new problems.

Version 13: “What if we fill the entire stone with the darkest shade first, then overlay lighter polygons on top?” This actually worked for the geometry problem — no more gaps. But now we had a different issue: the colors looked wrong. Coherent lighting across five irregularly-shaped stones at different angles is genuinely hard. The AI kept making stone 5 look like it was lit from below while stone 3 was lit from the left.

The gradient experiment: We tried linear SVG gradients instead of flat polygons. One gradient per stone, light-to-dark. It took 30 seconds to see this was a dead end. The gradients made every stone look like a gel button from 2005.

The computed approach: I had the AI write a Node.js script that would generate the SVG programmatically — stone outlines as coordinate arrays, brightness values per face, a function to map brightness to emerald shades. The idea was solid: separate the data from the rendering, iterate faster by tweaking numbers instead of hand-editing SVG. The script actually survived into the final version (it renders the production logo today). But the brightness-to-color mapping was inverted on the first run — every highlighted face came out dark and every shadow came out light. After fixing that, we were back to the same fundamental problem: mapping 3D lighting intuition to 2D polygon math. The script was the right tool. The shading approach was still wrong.

The Breakthrough: Stop Adding, Start Removing

After all those iterations, I was frustrated but not ready to give up on the concept. The stone shapes were right. The composition worked. It was just the rendering approach that was fighting us.

I went to Google and searched for stone illustrations — cairns, rock piles, geological drawings. Not because I wanted to copy something, but because nothing the AI had produced matched what I had in my head. Most results were either photorealistic (not useful for SVG) or too cartoonish (not the vibe). But a few ink-style illustrations caught my eye. Simple. Bold. Two tones.

That was the turning point. Instead of simulating 3D with multiple shades and computed lighting, just embrace the flatness. Emerald fill. Dark outlines. A few light emerald crescents for highlights on the upper-left faces. Thin detail lines for texture. That’s it.

The AI nailed this style in two iterations. Not fifteen. Two.

The lesson was obvious in hindsight: we’d been fighting the medium. SVG isn’t a 3D renderer. It’s not even a 2D painting tool. It’s a precision geometry format. The approaches that work in SVG are the ones that lean into precision — clean shapes, bold strokes, deliberate simplicity. The moment we stopped trying to fake something SVG can’t do and embraced what it does well, everything clicked.

What AI Actually Struggles With in SVG

This experience crystallized something I’d been sensing across many AI-built visual assets: there’s a specific gap between spatial reasoning and coordinate production that current AI models haven’t bridged.

The AI could discuss composition fluently. It understood “make it more angular” and “the shadow is on the wrong side.” It could iterate on structure all day — adding stones, changing proportions, adjusting the arch. And once we’d defined the two-tone ink style, it executed it cleanly. The reasoning was never the problem.

The problem was translating spatial intuition into numbers. “This face should be lighter because it catches sunlight” is a concept. Turning that concept into the eight polygon vertices that create the correct visual impression at the right coordinates, without overlapping adjacent polygons or leaving gaps — that’s a fundamentally different skill. The AI would produce coordinates that were plausible but off by 10 or 20 units. In the data, you can’t see the error. In the render, it’s immediately obvious.

But the deepest limitation wasn’t coordinate accuracy. It was choosing the right level of abstraction. The AI never once suggested “maybe we should simplify the approach.” It would happily keep generating increasingly complex shading systems, adding more polygons, more color stops, more brightness calculations — each one technically reasonable, none of them actually solving the problem.

The visual refiner that wouldn’t stop was a pattern I’d seen before in multi-agent orchestration. Agents caught in refinement loops, making things more complex instead of stepping back. The logo redesign was the same dynamic, just with a human in the loop. I was the one who had to say “this approach fundamentally won’t work” — and even I waited too long to say it.

The Final Logo

Five angular stones. An arch formed by the capstone leaning on the pillar. Dark emerald outlines thick enough to show the gaps between stones even at favicon size. Light emerald highlights on three stones suggesting directional light. Detail lines that add texture when you zoom in but disappear gracefully at small scales.

It’s not photorealistic. It doesn’t try to be. It’s a mark — like the cairns it represents. Clear enough to recognize, simple enough to scale, meaningful enough to remember.

Rocks, Not Robots

An AI consulting brand with a stone pile for a logo. It sounds like a contradiction.

But every neural network icon and circuit board logo is saying the same thing: “We are technology.” The cairn says something different: “We build things that last.” Stones stacked by travelers who came before, marking the path for those who come after. That’s the original spark behind this whole project — not just building with AI, but building systems that accumulate knowledge across time.

Compare this to rebranding a client’s color scheme, where the AI executed a full visual overhaul autonomously in 20 minutes. The difference? Color replacement is mechanical — find hex code, replace hex code. Logo design is spatial and aesthetic. The AI can do the first without human intervention. For the second, it needed 15+ rounds of human judgment to land on something that works.

That contrast — and the logo itself being built through those iterations, complete with failures, dead ends, and a crucial moment where the human had to override the AI’s approach — feels like the most honest brand statement I could make.

This is what agentic engineering actually looks like. Not magic. Not autonomous perfection. A human and an AI, stacking stones together, each one a little better placed than the last.