Why not capture this feedback as rules for the agent to not make the same mistakes?

You can, and it's a natural instinct. But if every piece of last mile feedback becomes a new rule, you've recreated the exact problem we spent years fighting in design systems. Without restraint, they spiral in both team size and complexity (more components, more states, more documentation). A pile of agent rules is just that same pile in a different file.

Most of this feedback is also specific. "Move this button up 8 pixels on the empty state of the settings screen" fixes one combination of layout and content. Written as a rule, it either applies too narrowly to matter or too broadly and breaks something else. And the longer the instructions get, the more they compete for the model's attention. Context windows (the amount of text a model can "remember" at once) help, but bigger ones only help to a point.

The useful work is spotting which fixes point to an underlying pattern and folding that into a simpler, more cohesive system. That takes judgment, and today's models don't have enough of it. Left alone, they'll add documentation instead of simplifying. Generating output keeps getting cheaper, but our capability is still capped by our ability to evaluate and synthesize.

I see the same thing with this Ask LukeW system. The real work is constantly reviewing responses and course correcting, which is applying taste and a perspective, not saving every correction verbatim.

So my advice: collect the feedback, look at it in batches every so often, and have a designer distill it into fewer, better patterns. Treat feedback as raw material for the design system, and keep a designer in charge of deciding what actually goes into it.