Because agents optimize for the average.
AI has made it easier than ever to ship software. A small team can build in days what used to take months. That part works.
What doesn’t work is that most of what gets built looks the same. Same layouts, same type, same patterns, same voice. Not because founders don’t have vision, because coding agents don’t have taste. They reach for the most likely answer, which is usually the average one. Good enough to ship. Not good enough to remember.
The teams we work with are moving fast and building directly in code, shipping features in Cursor, Claude Code, or whatever’s in their stack. The question is whether the things they ship actually sound like them; whether it holds together, stays coherent, and feels like it came from one mind instead of a hundred individual AI decisions. That gap is where we work.
We help define what a product should feel like before it gets built, then turn that into something your agents will actually follow. Not just a visual system. Three of them: one for how your product looks, one for how it sounds, and one for what it says. The character, voice, terminology, and the way the product handles errors and empty states and edge cases. Everything wires into your existing workflow: your CLAUDE.md, your rules, your repo. We extend what you have, not replace it.
Those decisions have always been easy to skip, and they’re even easier now. The tricky part is that even when they exist, agents drift. They fall back to defaults over time. A page here, a modal there, a new flow that’s almost right but not quite. It’s subtle at first and then it isn’t. The product starts to feel like it was built by committee, and no longer feels like it came from one mind.
Design systems for AI agents. Not a document you download and forget, but a live system your agents query every time they ship. That’s what we’re here to build.
- Brian & Noah
