Never delegate the understanding
Every year at WeAreDevelopers, hundreds of engineers share what they know on our stages. The recordings end up on YouTube, but the insights get lost: a 45-minute talk holds five minutes for you, five more for the next, and they stay buried. Last May, Thomas and I started Labs with one experiment: breaking long talks into moments. It grew into one search across everything we produce, down to the minute. There was a second reason. Rather than debate what AI means for us, we’d find out in what we build and how we build it.
I use agents for code, design, planning: almost everything. 18 months ago I lived in design tools; today I live in the terminal. Vibe coder in residence, if you like. It went in steps: ChatGPT, then Copilot, then one agent, now several at once. Every step cut another handoff between idea and code. Thomas and I have built products together for almost twenty years, but never this fast. Two people, no committee, shipping almost daily, a thousand-plus commits, a production-ready product, not a prototype. By November it felt like superpowers.
Superpowers made me sloppy. The easier it got, the more I told agents to “just do it”. Every just-do-it cost a bit of understanding, and more honestly, a bit of care. Quality slipped, sometimes we caught it, sometimes we didn’t, and that’s with four experienced eyeballs. It compounds: an agent bolts on a hundred lines where a rethink needs twenty. Speed pulled me out of my lane. I can judge design, UX, frontend, the shape of a system. The backend code, less so, and with agents I happily wandered in anyway. It worked, drifted, worked again. The cleanup was constant. I reviewed his PRs, he reviewed mine, and I still sometimes lost track. Approving code is not the same as understanding it.
So we slowed down, to be fast again later. That discipline, more than the agents, made it production-ready. Writing code is cheap now. Knowing what’s worth building isn’t, and neither is a teammate who cares whether the system makes sense. We’re on a call all day, taking each other’s work apart. That’s not code review, that’s an argument, and no document can give you one. The agents wrote our documentation, changelogs, design docs, release notes. The one thing that never wrote itself was my understanding.
A year in, I love what we built. And still: AI didn’t make me worse at my job, it made it easier to be worse without noticing. Bad code fails loudly; bad UX and lost understanding fail quietly, then all at once. Charles Eames built a hundred prototypes to get one chair right, and asked what he’d delegate, he said: never the understanding. That line has stuck with me since I studied architecture. AI is the first thing strong enough to make me nearly break it.