Something I’ve noticed as I rely more and more on fully autonomous agents to write code: it accelerates the rate at which I have to make harder decisions, introducing a new kind of bottleneck or fatigue.
Where this has become most obvious is in the design pressure that building things has put on me. As a UX designer, and someone who has designed countless products, generally it’s relatively easy for me to come up with a design for a product. In the past, while prototyping (by hand) or wireframing, I’d work through the hard parts of the design pretty methodically.
What I’ve noticed now is that these parts aren’t coming as easily. This is a little alarming - it’s because this is generally easy for me that I’ve been able to make a career out of it. So what gives?
I’ve got a few theories:
- Because AI increases my velocity so much, I end up parallelizing more and more projects, working on way more in a single day than I used to. It might be that because of that, I’m not allowing my brain to sit and run background cycles on a single problem, as I’m overloading it by multiplexing so much.
- Writing code by hand is a kind of natural limiter in terms of speed. You’re physically limited to the number of characters you can type in a minute, and cognitively limited to how much of the code and architecture you can effectively keep in your head at any given time. Because of this, the slowed pace naturally permits more thinking time to happen on other parts of the problem, namely design, which is being solved partly as a background thread while working on the code. With that natural limiter now removed, the background thread for design doesn’t have time to complete, and becomes a blocker.
- The baseline level of UX and UI design by agents now is surprisingly good. It’s not great, but it’s certainly raised the floor on what would be expected now in terms of general design quality. That means that designing something that stands out and gives a more unique experience is even harder, we have to up our game more in order to stand out from an increasingly average across the board.
Right now, I’m most tempted to say that the problem is the acceleration of the production time, and shortchanging the mental background thread that’s working on the design. Given that the resources are increasingly multiplexed as well, there’s even less cognitive energy spent on that task.
Not sure what to do about it, and maybe it’s just that the LLMs will get good enough at design that this won’t matter ultimately, but it’s an interesting observation of a side effect that I didn’t expect from working at this speed.