Hi Luke, I just read “Design Feedback for AI Agents.” I’ve been working out an approach to this problem myself, and I’m having a hard time finding others exploring this at this depth. If you know of anyone, I’d love to know who they are. I’ve spent a long time in large organizations, where there’s always a tension between “hurry up and build something” and "we have quality concerns." And this has become a real problem as orgs bring AI agents into the workflow. Inside my org I’ve been arguing that these tools give us a chance to rethink how we do UX design. We can use agents to understand the domain problem more deeply. We can also use them to track how a product experience and its design unfolded over time, which has always been possible but very time-consuming. Your point that the design system and instructions get a first pass about 90% of the way there, and that the rest comes from showing the agent where things don’t work, matches how I think about it. Understanding a design seems to depend on seeing it first. Something can be known without being understood, and the understanding arrives when you look at it and respond. I've been referring to it as a visual spike, to understand optionality. But the opportunity with the agents is to have those options presented with the agent understanding what exists, and the tradeoffs and tensions that will arise with one decision vs the other. I’m worried about how many organizations will skip this step, so I’ve committed to writing and publishing publicly on it, to help people outside the place I work at. Thanks for writing this up. I hope you have a good week. Cheers, Jason
Thanks Jason, I appreciate you reading it. I wish I had a good list of names for you, but I haven't written about specific people exploring this at the depth you're describing, so I can't point you to anyone with confidence. Which is a pretty good sign there's room for you to publish on it.
Your "visual spike" framing fits what we keep seeing. If you give today's models free license, they'll err on the side of over documentation instead of the pattern simplification and cohesion an effective design system needs. So the written system gets you most of the way, and the rest comes from looking at real output and reacting to it. The upside is that agents can read design system documentation every single time they write code, something design teams spent years trying to get people to do.
On the "hurry up and build" versus quality tension: a lot of what I see right now is people saying they don't need designers or developers anymore because they made a prototype in five minutes. But very few are looking at all these capabilities and asking how we actually work together with them. Ten fast prototypes in a meeting isn't a process. Here's roughly how we've been sketching that collaboration out, with designers, PMs, and developers all building in parallel and developers reviewing:
Your idea of having the agent present options along with the tradeoffs of each is a good extension of this. It moves the agent from producing a single answer to helping you see the decision space, which is where design judgment actually gets applied.
As for orgs skipping this step: I suspect many will, at least at first. I've always held that the most important role of designers is fighting for the coherency, simplicity, and visual communication needed to humanize technology. New tools don't change this. New ways of working together do. Writing publicly about how you're doing it is one of the best ways to help other teams figure that out, so keep going.
Have a good week too.