Reading through everything the Design Systems Collective community published in June, the thing that struck me first wasn’t agreement. It was how loudly the community disagreed with itself, about consistency, about tokens, about whether a design system is even the right thing to be building, while arriving at the same anxious question from every direction: does the system know what it actually is.
This month’s newsletter kept returning to one strand of that question, the AI-agent one: whether the system you documented is the system you have, whether taste survives being reduced to a token, whether a system that looks fine to a human reviewer is quietly failing an agent, whether the agent is now an author rather than just a reader, whether governance is architecture rather than a meeting. Ninety of June’s other articles either back that arc with sharper evidence than I had room for, or they were having an entirely different argument I hadn’t made room for at all. Both matter. Here’s both.
The Gap Nobody Documented, Now With Receipts
The strand I’d already been writing about got sharper this month, not because the idea changed but because the community started producing evidence for it instead of just theorising. Murphy Trueman’s piece on the parts of your system you never wrote down is the clearest version I read all month: an AI agent doesn’t ask a clarifying question when it hits an undocumented decision, it fills the gap with the average pattern from everywhere else it’s been trained, and nobody notices because the output looks plausible. Surendar Selvaraj made the same point twice from different angles, once arguing your design system now has two audiences and you only documented it for one, and once with a name for the fix, a Dual Contribution Model where human and machine contributions get reviewed against different standards because they fail in different ways.
Eddie Lou’s closing piece of the month put the sharpest possible frame on it: AI agents are a fundamentally different consumer of your system than a human ever was, because a confused human slows down and asks someone. A confused agent doesn’t slow down. It just scales the wrong answer, silently, until someone finally goes looking for why the product feels like it’s drifting.
That’s not a new failure mode. It’s an old one running at a speed nobody built instrumentation for.
Design System, Or Just a Nicer Word for Kit
The part I hadn’t made room for in the newsletters was the argument happening one layer down, about whether the thing being defended is even a design system in the first place. Matilda Anashie’s piece, bluntly titled Stop Calling It a Design System, You Just Want a Prettier Style Guide, and Carlos Fraccalvieri’s contrarian You Don’t Need a Design System, You Need Design Discipline, are both making the same accusation from opposite ends: most of what gets called a design system is a component kit with a shared palette and a governance story nobody’s actually enforcing. Nadiia Abrosymova’s framing of the artifact layer versus the pattern layer is the most useful version of this I read all month. AI tools can extract the artifact layer, the tokens and the components, in an afternoon. What they can’t extract is the pattern layer, the unwritten reasons those artifacts exist, because most teams never wrote the pattern layer down in the first place. That’s not a new critique. It’s the same one design systems people have been making about style guides for a decade. It’s just true again, faster, because AI is now the thing exposing it.
Underneath that sits a genuine disagreement I don’t think the community has resolved, and I’m not sure it should pretend to. Leo Lopes wants tokens treated as hard, CI-enforced constraints that fail the build the moment someone deviates, and George William Amalan goes further, arguing most teams are carrying AI debt they haven’t even measured yet. Surendar Selvaraj, watching Apple and Google rewrite their own platform tokens mid-year, argues the opposite: tokens are becoming policies, ranges, not fixed values, because rigid enforcement is exactly what breaks first when the platform itself changes its mind. Shadi Abd’s Hidden Cost of Over-Consistency sides with flexibility, arguing enterprise systems routinely mistake uniformity for consistency and punish usability for it. My honest read: the CI-enforcement camp is solving for AI governance, the flexibility camp is solving for platform reality and nobody’s written the piece that reconciles the two yet. Someone should.
Consistency Solved One Problem and Started Another
The other argument I’d left out entirely was about character. IAMJAMES wrote three pieces this month essentially building one thesis: design systems professionalised the web and, in doing so, solved consistency and lost character. The result is an internet that’s becoming visually exhausting and a web where most websites don’t need more features, they need better taste, because everything got optimised for correctness instead of quality. Rakesh patel’s piece, arguing that every great experience feels familiar because predictability builds trust, isn’t wrong so much as answering a different question. Familiarity and character aren’t actually opposites, but the community wrote about them all month as though they were, and I think that’s worth naming rather than smoothing over.
Three separate writers also landed on the same reference point within days of each other. Abhi Chatterjee’s take on Figma Config 2026, Surendar Selvaraj’s From Design Files to Living Product Systems and Madhesh P’s read on what actually matters out of Config all converged on the same line without coordinating: AI lowered the floor, and the ceiling is now the only thing left to compete on.
That line traces straight back to Dylan Field. At the keynote, Figma’s co-founder and CEO put it plainly:
“At past Configs, we’ve talked about AI lowering the floor and raising the ceiling. But while AI has lowered the floor, it has not raised the ceiling. That part is on us.”
The community read that as a challenge rather than a slogan. AI has democratised and accelerated the early stages of design, but the burden of real innovation still sits with designers, pushing through collaborative workflows, bringing code closer to design and using the new materials Figma introduced this year, motion and shaders, to actually raise the bar instead of waiting for AI to raise it for them.
What I Think Is Actually True
Here’s where I land, having read all of it. The AI-agent infrastructure argument I’ve been making since May is correct, and June’s corpus backs it with sharper evidence than I had a month ago. But it isn’t the whole story, and the pieces I’d have missed if I only read my own newsletters back are the ones asking whether the thing being made AI-legible was ever worth reproducing at scale. A design system that’s rigorously tokenised, CI-enforced and perfectly machine-readable can still be a style guide with better manners. Making the kit legible to a machine was never going to fix the fact that, for a lot of teams, nobody wrote the judgment down because nobody agreed on it in the first place.
The community spent June arguing with itself about consistency, about enforcement, about what even counts as a system. I don’t think that argument is a sign the field is confused. I think it’s a sign the field is finally being honest about a distinction it’s spent a decade blurring: between a system that looks coherent and one that actually is.